"Logics" by John Eric Nolt
"Logics" by John Eric Nolt
"Logics" by John Eric Nolt
To get extro volue from this book for no odditionol cost, go to:
JOHN NOLT
U niu er sity of Tenne ss e e, Kno xuille
͡
/ ヽ
(11】 :】:ア
V
リ
Wadsryorth Publishing Company
I(DP' An International Thomson Publishing Company
Belmont, CA o Albany, NY o Bonn . Boston o Cincinnati . Detroir o Johannesburg
London r Madrid r Melbourne o Mexico City o 11.* York o Paris o San Francisco
Singapore . Tokyo o Toronto o \Washington
Philosophy Editor: Peter Adams Advertising Project Manager: Joseph Jodar
Assistant Editor: Clayton Glad Designer: John Edeen
Editorial Assistant: Greg Bruek Copy Editor: Margaret Moore
Project Editor: Gary Mcdonald Cover Designer: Randall Goodall
Production: Greg Hubit Bookworks Compositor: Thompson Type
Print Buyer: Barbara Britton Printer: Quebecor Printing / Fairfield
E)u:xbury Press and the leaf1030 are tradcn■ arks used undcr license.
Printcd in the unitcd States Of Amcrica
12345678910
All rights reserved. No part of this work covered by the copyright hereon may be
reproduced or used in any form or by any means-graphic, electronic, or mechanical,
including photocopying, recording, taping, or information storage and retrieval systems-
without the written permission of the publisher.
Preface xi
V‖
viii Conre rurs
:ndex 463
PREFACE
This book has many uses. Chapte rs 7,2, 3, 4, 6,7, and 8 provide the basics for a
one-term introduction to formal logic. But later chapters contain an ample stock
of advanced material as well, allowing for a variety of two-course sequences. It
may also do duty for a second course alone, in which the early chapters provide
review and the later chapters a selection of topics.
Regardless of how it is used, this book is designed to meet several specific
needs. There is, first of all, the need to convey to students some of the diversity of
recent developments in logic. Logic, as the title is intended to suggest, is no longer
just logic but logics-a study of a multitude of systems developed for an impressive
variety of applications. Though classical predicate logic is still the centerpiece, it is
by no means all, or even most, of the subiect. From the beginning, this book makes
the presuppositions of classical logic explicit and points to alternatives.
Second, this is a text that seeks to balance the formal and philosophical with
the practical. A wide range of formal topics is covered, and there is frequent
reference to their philosophical roots. But in no case do I treat any system as merely
a formal object. Logic is, first and foremost, the study of reasoning, and its life-
blood is the elucidation of particular arguments. Thus, even where this book ex-
amines exotic formal systems, practical understanding of inference is always the
primary concern.
Third, to facilitate understanding, each system is introduced, first, by way of
concrete problems that motivate it and then by an account of its semantics. Proof
theory, though usually historically prior, is relegated to third place, since much
that is puzzling about proofs can be elucidated semantically, whereas relatively
little that is puzzling about semantics can be illuminated by proofs. The ultimate
step for each system is an ascent to the vantage point of metatheory, where the
deepest understanding may be achieved.
In doing semantics, some metatheory is, of course, unavoidable. The main
issue is how explicit to be about it. I have been very explicit. Metatheory baffles
many students chiefly because the rules of the game are rarely explained. In the
first five sections of Chapter 5 I have endeavored to explain them'
With respect to metatheory itself, my aim has been to err on the side of too
much help, rather than not enough. I hope, however, that that aim is not incom-
patible with elegance. Detailed explanations generally precede the more difficult
metaproofs, but the metaproofs themselves are as simple and nontechnical as I can
make them.
xii Pnrracr
Logic is the study of reasoning. Reasoning is a process ofthought, but there exists
no uncontroversial method for studying thought. As a result, contemporary logic,
which likes to think of itself as founded on hard (though perhaps not empirical)
facts, has nothing to say about thought. Instead, logicians study certain excres-
cences of thought: verbalized bits of reasoning-arguments'
An argument is a sequence of declarative sentences, one of which is intended
as a conclusion; the remaining sentences, the premises, are intended to prove or
at least provide some evidence for the conclusion. The premises and conclusion
express propositions-which may be true or false-as opposed to questions, com-
mands, or exclamations. Nondeclarative Sentences may sometimes Suggest prem-
ises or conclusions, but they rlever Are premises or conclusions.
Declarative sentences are not themselves propositions. Some theorists have
held that propositions are assertions made by sentences in particular contexts;
others, that they are the meanings of sentences or the thoughts sentences express.
But it is generally agreed that between sentences and propositions there is an
important difference. The sentence "I am a woman" uttered by me expresses a
difierent proposition than the same sentence uttered by you. \When I utter the
sentence, ihe proposltion I assert is false; if you are a woman, when you utter the
3
4 CHaprrn 1
sentence you assert a true proposition. Even if you are not a woman, the proposi-
tion you assert by uttering this sentence is different than the one I assert by uttering
it; your proposition is about you, mine about me.
Logicians, however, tend in practice to ignore the differences between sen-
tences and propositions, studying the former as if they were the latter. This practice
presupposes that each argument we study is uttered in a fixed context (a given
speaker in a given circumstance), since only relative to such a fixed context does
each sentence in the argument express a unique proposition. To illustrate, consider
the following argument:
itself. Still, bad as it is, it's an argument; the author intended the first two proposi-
tions (sentences) to be taken as evidence for or proof of the third, and that's all
that being an argument requires.
Let's now consider a good one. I'll begin with a claim. The claim is that in
certain matters your will is not free. In fact there is one act you cannot initiate no
matter how strong your will. The act is this: to criticize all and only those people
who are un-self-critical. For example, consider yourself. Are you going to criticize
yourself or not? If you do, then you will criticize someone who is self-critical
(namel5 you)-and so you're not criticizing only the un-self-critical. On the other
hand, if you don't criticize yourself, then you fail to criticize someone who is un-
self-critical (namely, you again)-and so you don't criticize all the un-self-critical.
So either way you fail.'
Now consider your thoughts as you read the previous paragraph. (I assume
you read it with comprehension; if not, now might be a good time to try again.)
When I first made the claim, unless you had read this sort of thing before, you
were probably puzzled. You wondered, among other things, what I was up to. At
a cer;in point (or maybe not a certain point-maybe slowly), a light went on and
you saw it. The dawning of that light is insight.
A good argument, when it works, gives you insight. It enables you to see
why the conclusion is true-not "see" in a literal sense, of course, but "in your
mind's eye." !7hat was wrong with the bad argument given above was that it
didn't yield any insight at all. It puzzled us and offered no resolution to our
puzzlement.
Here I am talking about thought (insight, puzzlement, dawning lights, and
so on), when I said just a few paragraphs back that we were going to talk about
symbol systems. That's because I want to make vivid a certain contrast. There is
much to be noticed about the experience-the phenomenology-of argumenta-
tion. But contemporary logicians try to explain as much as possible of what makes
an argument good or bad without using mentalistic iargon, which they view with
suspicion. They prefer to talk about symbols.
The previous argument showed us something about insight, but it's rather
flashy for an introductory illustrationl let's consider a more mundane and time-
worn example:
All men are mortal.
Socrates is a man.
Socrates is mortal.2
1 The reasoning here is identical to the reasoning of Russell's barber paradox and to
the core of the argument by which we will prove the halting problem unsolvable
(see Section 10.5).
2 The origin of this argument is a mystery to me. It appears in many logic text-
books, going way back in history, so presumably it has a classical source' The
obvious source would be Aristotle, since Aristotle invented formal logic, but an
Aristotle scholar assures me that this argument is nowhere to be found among the
Philosopher's works.
5 Cuapren 1
This is good too, if not so good as our last example. It could, I suppose, convey
some insight to a sheltered three-year-old. Its virtues, according to hoary tradition,
are these:
That leaves us with virtue 2, the one that generally interests logicians. Most logi-
cians have belonged to a school of thought known as the classical tradition. In the
first four parts of this book we will consider logic from the classical perspective,
though in the fifth we shall step outside of it. To say rhar an argument is valid is,
according to the classical tradition, to say that there is no way for the conclusion
not to be true while the premises are true. We'll sometimes put this in terms of
"possible situations": There is no possible situation in which the premises are true
but the conclusion isn't.
The Socrates argument is valid, for there is no possible situation in which all
men are mortal, Socrates is a man, and Socrates is not mortal; we can't even
coherently think such a thing.
The end-in-itself argument is invalid (i.e., not valid), for there is a possible
situation in which the premises are rrue and the conclusion isn't. That is, it is
possible that humans are the only rational beings and that rationality alone en-
ables a being to make moral judgments but that humans are not the only ends-in-
themselves. One way this is possible is if being an end-in-itself has nothing to do
with the ability to make moral judgments, but rather is linked to some more
general capacity, such as sentience or the ability to live and flourish. Thus perhaps
other critters are also ends-in-themselves even if the argument's premises are true.
A possible situation in which an argument's premises are true and its conclu-
sion is not true is called a counterexample to the argument. \07e may define validity
more briefly simply by saying that a valid argument is one without a
counterexample.
'When
we speak of possible situations, the term 'possible' is to be understood
in a very broad sense. To be possible, a situation need not be something we can
lruronr,.aal Loctc 7
bring aboutl it doesn't even have to obey the laws of physics. It just has to be
something we can coherently conceive-that is, it has to be thinkable and describ-
able without self-contradiction.
Thus, intuitively,3 to tell whether or not an argument is valid, we try to
conceive or imagine a possible situation in which its premises are true and conclu-
sion is untrue. If we succeed (i.e., if we can describe a counterexample), the argu-
ment is invalid. If we fail, then either we have not been imaginative enough or the
argument is valid. This makes logicians neryous; they'd like to have a test that
doesn't rely on human ingenuity; much of this book will be devoted to explaining
what they do about this anxiety and how their efforts fare.
'We
But most people are not so skittish. appeal to counterexamples almost
unconsciously in everyday life. Consider this mundane argument:
They said on the radio that it's going to be a beautiful day today.
It is going to be beautiful today.
One natural (albeit cynical) reply is, "They could be wrong." This reply demon-
strates the invalidity of the argument by describing a counterexample-that is, a
possible situation in which the conclusion ('It's going to be a beautiful day today')
is untrue even though the premise ('They said so on the radio') is true: namelS the
situation in which the forecasters are wrong.
A counterexample need not be an actual situation, though it might; it is
enough that the situation be conceptually possible. Thus it need not be true that
the forecasters are wrong; to see the invalidity of the argument, we need only
realize that it is possible they are wrong.
To give a counterexample, then, is merely to tell a kind of story. The story
needn't be true, but it must be conceptually coherent. The cynical respondent to
our argument above hints at such a story with the remark "They could be wrong."
That's enough for casual conversation. But for logical analysis it's useful to
be more explicit. A well-stated description of a counterexample should contain
three elements:
If we flesh out the cynic's counterexample to make all of these elements explicit,
the result might be something like this:
They said on the radio that it's going to be a beautiful day today. But they
are wrong. A cold front is moving in unexpectedly and will bring rain in-
stead of a beautiful day.
3 r0(hen I say 'intuitively', I mean from an informal point of view. We are still talk
ing about thoughts here, not symbols. This is typical of informal logic. The for-
mal, symbolic approach begins with the next chapter.
8 Cuapren I
All three elements are now present. The first sentence of this "story" affirms the
premise. The second denies the conclusion. The third explains how the conclusion
could be untrue even though the premise is true.
This is not, of course, the only possible situation that would make the prem-
ises but not the conclusion true. I made up the idea of an unexpected cold front
more or less arbitrarily. There are other counterexamples as well. Maybe an un-
expected warm front will bring rain. Or maybe there will be an unexpected dust
storm. Or maybe the radio announcer knew it was going to be an awful day and
flat out lied. Each of these scenarios is a counterexample. This is typical; invalid
arguments usually have indefinitely many counterexamples, each of which is by
itself sufficient to show that the argument is invalid.
Let's consider another example. Is the following argument valid or invalid?
Again all three elements of a well-described counterexample are present. The state-
ment'All philosophers are freethinkers and Al is not a philosopher' affirms both
of the premises. The statement 'Al is nevertheless a freethinker' denies the conclu-
sion, and the remainder of the story explains how this can be so. The story is
perfectly coherent, and thus it shows us how the conclusion could be untrue even
if the premises were true.
Notice again that the counterexample need not be an actual situation. It's
just a story, a scenario, a fiction. In fact, it isn't true that all philosophers are
freethinkers, and maybe it isn't true thar Al (whoever Al is) is a freethinker, either.
That doesn't matter; our story still provides a counterexample, and it shows that
the argument is invalid, by showing how it cowld be that the conclusion is untrue
while the premises are true.
Notice, further, that we needn't have said that Al is a bricklayer; for purposes
of the example, he could have been an anarcho-communist or some other species
of freethinker-or an unspecified kind of freethinker. The details are flexible; what
counts, however we formulate the details, is that our "story" is coherent and that
it makes the premises true and the conclusion untrue.
Let's consider another argument:
This story does seem to make the premises of the argument true and the conclusion
false. But is it really intelligible? If having an immortal soul makes one immortal
and the man Socrates has an immortal soul, then not all men are mortal. The story
is incoherent; it contradicts itself. It is therefore not a genuine counterexample,
since a counterexample is a possible situation; that is, its description must be
conceptually coherent.
Some additional invalid arguments with accompanying counterexamples are
listed below. Keep in mind that invalid arguments generally have many counter-
examples so that the counterexamples presented here are not the only ones. Note
also that each counterexample contains all three elements (though sometimes more
than one element may be expressed by the same sentence). The three elements,
once again, are
!nvalid Argument
Sandy is not a man.
Sandy is a woman.
Counterexample
Sandy is neither a man nor a woman but a hamster.
10 CHrpren l
lnvalid Argument
If the TV is unplugged, it doesn't work.
The TV is not working.
It's unplugged.
Counterexample
If the TV is unplugged it doesn't work, and it's not working. However,
it is plugged in. The reason it's not working is that there's a short in the
circuitry.
lnvalid Argument
All charged particles have mass.
Neutrons are particles that have mass.
Neutrons are charged particles.
Counterexample
All charged particles have mass, but so do some uncharged particles,
including neutrons.
lnvalid Argument
The winning ticket is number 540.
Beth holds ticket number 539.
Beth does not hold the winning ticket.
Counterexample
The winning ticket is number 540; Beth is holding both ticket 539 and
ticket 540.
lnvalid Argument
There is nobody in this room taller than Amy.
Bill is in this room.
Bill is shorter than Amy.
Counterexample
Bill and Amy are the only ones in this room, and they are the same
height.
lruronvar Loctc 11
lnvalid Argument
Sally does not believe that Eve ate the apple,
Sally believes that Eve did not eat the apple.
Counterexample
Sally has no opinion about the story of Eve. She doesn't believe that Eve
ate the apple, but she doesn't disbelieve it either.
lnvalid Argument
Some people smoke cigars.
Some people smoke pipes.
Some people smoke both cigars and pipes.
Counterexample
There are pipe-smokers and cigar-smokers, but nobody smokes both
pipes and cigars, so the two groups don't have any members in common.
lnvalid Argument
Some people smoke cigars.
Some people do not smoke cigars.
Counterexample
There are people, and all of them smoke cigars. (If everybody does, then
some people do and so the premise is true!)
lnvalid Argument
We need to raise some money for our club.
Having a bake sale would raise money.
'We
should have a bake sale.
Counterexample
'We
need to raise money for the club, and having a bake sale would raise
money, but so would other kinds of events, like holding a car wash or a
telethon. Some of these alternative fund-raising ideas better suit the
needs of the club and the abilities of its members, and so they are what
should be done instead of a bake sale.
12 CHnprEn 1
lnvalid Argument
Kate hit me first.
I had to hit her back.
Counterexample
Kate hit the (obviously immature) arguer first. But the arguer could have
turned the other cheek or simply walked away; there was no need to hit
back.
'!(hat launched
Let's take stock. our discussion of counterexamples was talk
of validity, and what led us to validity was a look at the two virtues of a good
argument, namely:
1. The premises are true.
2. The reasoning is valid.
Logicians sometimes suggest that these two virtues are sufficient for a good argu-
ment. I have already expressed doubts about this. But we can see why someone
might believe it if we consider the two virtues together. To say that the reasoning
is valid is to say that there is no counterexample-that is, there is no way for the
conclusion not to be true while the premises are true. Now, if we add virtue 1-
namelS that the premises are true_we see that the two virtues together add up to
a guarantee of the truth of the conclusion. An argument that has both virtues-
true premises and valid reasoning-is said to be sound. Sound reasoning certifies
that its conclusion is true.
If that's all we want from reasoning, then virtues I and 2 are all we need. In
the classical logical tradition, it has been customary to ask for no more. But I think
.We
we generally want more. want insight, significance, cogency . . . well, at least
we want relevance. Virtues 1. and 2 don't even give us that-as we shall see in the
next section.
Exercise 1.2
Classify the following arguments as valid or invalid. For those rhat are invalid,
describe a counterexample, making sure that your description includes all three
elements of a well-described counterexample. Take each argument as it stands;
that is, don't alter the problem by, for example, adding premises.
1. No plants are sentient.
All morally considerable things are sentienr.
No plants are morally considerable.
2. All mathematical truths are knowable.
All mathematical truths are eternal.
All that is knowable is eternal.
3. Most geniuses have been close to madness.
Blake was a genius.
Blake was close to madness.
lruronvar Locrc 13
4 . 5 . 6 .
The surface area of Russia is larger than the surface area of China.
9.
The witnesses said that either one or two shots were fired at the victim.
Two bullets were found in the victim's body.
Two shots were fired at the victim.
Lt. People do climb Mount Everest without oxygen tanks.
It is possible to climb Mount Everest without oxygen tanks.
12. Some fools are greedy.
Some fools are lecherous.
There are some fools who are both lecherous and greedy.
13. No one has ever lived for 200 years.
No one ever will.
14. DNA contains the code of life.
Life is sacred.
It is wrong to manipulate DNA.
15. There are fewer than a billion people in the whole United States.
New York is only a part of the United States.
There aren't a billion people in New York.
1.3 RELEVANCE
Prettybad-but it has both of the virtues discussed in the previous section: The
premise is true, and the reasoning is valid. Of course the conclusion doesn'tfollow
from the premise. But that wasn't how we defined validity-following from the
premises. \We defined it as the absence of a counterexample. And there is no coun-
terexample here.
14 CHaprrn 1
The queerness resides in the conclusion: 'There's no tree that's not a tree'.
This conclusion can't be untrue; it's true in any possible situation, no matter what
the world is like. Hence there is no possible situation in which the premise is true
and the conclusion is not (simply because, regardless of the premises, there is no
possible situation in which the conclusion is not true). So the argument has no
counterexample; it is valid. Further, since the premise is true, it is sound. Still, it is
a dumb argument.
Not that it leads to an incorrect conclusion. The conclusion can't be untrue;
it's true in all possible situations. So here as elsewhere, soundness guarantees truth.
What's wrong with this argument is that the conclusion derives no support from
the premise. The premise is irrelevant; the conclusion could stand on its own.
The conclusion is a logical truth, a statement true in all possible situations.
And the argument is an illustration of the general rule that any argument whose
conclusion is logically true is automatically ualid, no matter what the premises. A
logical truth must be true no matter what we assume; so it is a valid conclusion
from anything.
But since this argument is bad nevertheless, we may infer that at least one
additional virtue is required for good reasoning: relevance. But what is relevance?
In recent years a whole field of logic, relevance logic, has emerged to attempt to
answer this question. \We shall consider it in some detail in Section 15.3. Unfortu-
nately, there seem to be as many relevance logics as relevance logicians, so the
discipline is in disarray. But the need for relevance is clear.
There is another kind of inference in which the need for relevance stands out
starkly: an argument with inconsistent premises. A set of propositions (or a single
proposition) is inconsistent if there is no possible situation in which they are all
true (or in which it is true). The proposition
There is a tree that's not a tree.
Albert is a pirate.
Albert is not a pirate,
and the more complex set:
The fundamental ideas in the conclusion are 'course numbered less than 400'and
'being taken for graduate credit'. The first of these ideas comes from the first
premise and the second from the second. Each has its origin in a premise, and this
accounts, at least in part, for the conclusion's relevance.
Notice that I did not list the terms 'no' and 'can' as expressing ideas' These
terms represent logical relationships, and they belong to a class of words that we
shall call logical operators. (Some authors call them syncategorematic terms.)
RoughlS a word is a logical operator if it expresses, not a specific idea itself, but a
way of modifying or combining ideas. Some common logical operators are 'all',
tsome', tmost', tno', tnott, 'if
. . . then', 'or', 'unless', and 'and'. Thus, though a
relevant conclusion may not introduce ideas not contained in the premises, it may
use logical operators to recombine the premises' ideas in new ways. (Precisely
which such combinations preserve relevance is one of the controversial issues in
relevance logic.)
Contrast the previous relevant and valid argument with this argument'
which is both fallacious and irrelevant:
Smoking is harmful.
Smoking should be illegal.
15 CHaprun 1
There are two fundamental ideas in the conclusion-and perhaps a third. The two
obvious ones are 'smoking' and the notion of being 'illegal'. Depending on how
we count, we might also treat 'should' as an idea, though some logicians would
consider it a logical operator. No matter. The inference is clearly irrelevant, be-
cause although the term 'smoking' in the conclusion has its origin in the premise,
the term 'illegal' comes from nowhere-and that's a hallmark of irrelevance.
Notice that we could make the inference relevant by adding a premise con-
necting the idea of harm to the idea of illegality:
The conclusion is now relevantly drawn. The terms 'illegal' and 'should' (if we
want to count the latter as expressing an idea) come from the first premise, and
the term 'smoking' comes from the second. We have also strengthened the reason-
ing; the argument is now valid. The added premise, however, is false. And so,
though valid, the argument is unsound.
Arguments with logically true conclusions or inconsistent sets of premises
provide the most glaring examples of validity (in the classical sense) without rele-
vance. But logically true statements and inconsistent premise sets are relatively
rare in actual reasoning. Most of the statements with which we reason are
contingent-that is, true in some possible situations and false in others. !7hen we
reason with contingent statements, there is less dissonance between classical logic
and relevance logic. Still, differences remain, as we shall see in Section 16.3.
Exercise 1.3
Classify the following arguments as valid or invalid, using the informal concept of
validity. For those that are invalid, describe a counterexample. Then discuss
whether or not the argument's premises are relevant to its conclusion.
1.4 ARGUMENT:NDiCAITORS
the premise indicator 'because' signals that the statement 'it is indivisible' (where
'it' refers to the soul) is a premise supporting the conclusion 'the soul is indestruc-
tible'. Premise indicators can also occur at the beginnings of sentences, but the rule
still holds: The statement to which the premise indicator is attached is the premise;
the other is the conclusion. Hence, for example, in the sentence
the word 'since' shows that the statement'numbers are nonphysical'is a premise
leading to the conclusion'nonphysical objects exist'.
Conclusion indicators are words or phrases that signify that the statement to
which they are attached is a conclusion that follows from previously stated prem-
ises. English is rich in conclusion indicators. Some of the most common are 'there-
fore', 'thus', 'so', 'hence', 'then', 'it follows that', 'in conclusion', 'accordingly',
and 'consequently'. In the following argument, for example, 'hence' indicates that
the third statement, 'God exists', is a conclusion from the first two:
'!(ithout God, there can be no morality. Yet morality exists. Hence God
exists.
God exists, /or without God there can be no moralitS and morality exists.
These are three different expressions of the same argument. There are many
others. Notice that the conclusion (in this case, 'God exists') may occur at the end,
or at the beginning, or in the middle of the argument, depending on the arrange-
ment of argument indicators. All three positions are common in ordinary speech
and writing. But for logical analysis it is customary to list the premises first and
18 CHapruR 1
the conclusion, prefixed by '.'.', last, as we have been doing. This is called standard
form.
Arguments may also be stated without indicators, in which case we must rely
on subtler clues of context, intonation, or order to discern their structure. Most
often when argument indicators are lacking the conclusion is given first, followed
by the premises. Here is an example:
Ailanthus trees have smooth bark, but the bark of this tree is rough.
Exercise 1.4
Some of the following passages are arguments, some are not. Some of the argu-
ments are incomplete, lacking a premise, or a conclusion, or both. Rewrite each
argument in standard form, supplying implicit premises or conclusions. For those
passages that are not arguments, write 'not an argument'.
1. Uranium is heavier than iron, because gold is heavier than iron and uranium
is heavier than gold.
2. Since anyone under 18 is a juvenile and juveniles are not allowed on the
premises, Sally is not allowed on the premises.
3. If there is a storm warning, the siren sounds. So there is no storm warning,
since the siren is not sounding.
4. Savage could not have been the thief. The thief was over six feet tall. But
Savage is only 5'8 ".
5. I7e went to Indianapolis; then we went to Chicago.
6. The water froze, and when water freezes the temperature must be at or below
zero degrees Celsius.
7. Different culrures have different conceptions of rationality. Hence rationality
itself takes many forms, for what a culture conceiues as rational ts rational for
that culture.
8. Alice has a National Rifle Association sticker on her windshield. It is likelS
therefore, that she opposes gun control.
9. Because all things other than pleasure are valued only for the pleasure they
produce, but pleasure is valued for its own sake, only pleasure is intrinsically
valuable. For a thing is intrinsically valuable if and only if it is valued for its
own sake.
10. I lied because I was afraid you would hate me if I told the truth.
20 Csaprrn 1
Smog is not a word at all; it is a form of air pollution. In the first sentence the word
'smog' is mentioned but not used; in the second it is used but not mentioned. In
logic we usually mention specific symbols by using their quotation names in the
manner of the first senrence. This is why single quotation marks have appeared
and will continue to appear so frequently in this book.
Failure to observe the use/mention distinction can lead to confusion or non-
sense. Consider
rThich king? And why would he want to do that? But, of course, what is intended
is
-at
least until appropriate quotation marks are added, when it becomes this
simple and obvious truth:
thing else again. They are not shapes or marks and cannot themselves be written-
though they may be named. The names of the number ten are legion. They include
not only the Arabic numeral '10' but also the Roman numeral 'X', rhe formula '8
+ 2', the English word 'ten', and so on. Ten is not any of these things, but the
unique thing that they all name.
To summarize, when you see single quotation marks, look between them:
The word or phrase, symbol or formula that you see written there is, precisely as
written, the thing being mentioned. Where there are no quotation marks, the
words in the sentence are being used, and you generally have to look elsewhere to
find what is being mentioned.
'When
we are using one language to study another, the one we use is called
the metalanguage and the one being studied-i.e., mentioned-is called the object
language. For example, when native speakers of English study Hebrew, they usu-
ally converse about Hebrew grammar, style, wording, and so on in English. Here
English is the metalanguage and Hebrew is the object language. In the succeeding
chapters, we shall study various logical languages. Each, as we study it in turn,
will become our object language; but the metalanguage willalways be English. We
will, however, from time to time import exotica, such as Greek letters for variables,
or some notation from mathematics into our metalanguage, so it will be a special-
ized or technical form of English.
Exercise 1.5
One or more of the following sentences is true as it stands; others are not true
unless quotation marks are added. Supply quotation marKs where necessary to
make them true.
2.l ARGUMENTFORMS
In Chapter 1 we approached logic from an informal point of view, considering
arguments as they occur in natural language. Here we begin the study of formal
logic, whose subject matter is argument forms-patterns of reasoning shared by
many different arguments. Here is a simple argument form:
If P, then Q
P
a
This form is known by its medieval Latin name, tnodus ponens.l The letters 'P'
and 'Q' function as place-holders
.We
for declarative sentences. rUTe shall call such
letters sentence letters. say that an argument is an instance of a form comprised
of sentence letters (or, simply, that it bas that form) if it is obtainable from the
form by replacing the sentence letters with sentences, each occurrence of the same
25
26 Csaprrn 2
letter being replaced by the same sentence.2 The following argument, for example,
is an instance of the form modus ponens in which 'P' is replaced by 'the fetus is a
person' and'Q' by'abortion is murder':
Since the number of declarative sentences is potentially infinite, the form represents
infinitely many different arguments, all with the same structure. Another example
of the form modus ponens is
Iilfhat is significant about this form is that any argument that has it is valid. Since
there are infinitely many such arguments, to know that they all are valid is to
possess knowledge that is in a sense infinite. But can we really know that each
instance of modus ponens is valid?
One possible approach to such knowledge is to insert random sentences in
place of 'P' and 'Q' and check the resulting arguments for validity by trying to
formulate counterexamples. But this case-by-case approach, though it might ulti-
mately convince us, is logically inconclusive. Even in an entire human lifetime we
could check only a finite number of instances; there will always be infinitely many
that we have never examined. Maybe some of these, so strange and complex that
we would never think to check them, are invalid. Maybe even some of the ones we
have checked are invalid, and failures of imagination have prevented us from
noticing!
A more sophisticated approach is to move to a more abstract level of
thought. We might, for example, reason this way: No matter which sentences'P'
and 'Q' stand for, whenever 'if P, then Q' is true and 'P' is true, then 'Q' has to be
true as well. Here we focus on the general pattern of the reasoning, on the form,
rather than the form's instances. In this way we might be able to "see" that the
form itself guarantees the validity of its instances. But this sort of "seeing," or
intuition, is still fallible and hence still subject to doubt.
2 Notice, however, that this definition does not rule out replacing different letters
with the same sentences. Just as in the mathematical equation 'x + y = y + x' we
may legitimately replace both 'x' and 'y' by '2' to obtain the instance '2 + 2 = 2 +
2', so in the form modus ponens, for example, we might replace both'P'and 'Q'
by the same sentence, say'People have souls'. The result
is, like '2 + 2 = 2 + 2' , trivial and uninteresting. But, like all instances of modus
ponens, it is valid.
Crassrcal PRopostrtoNar Loctc: Svrurax 27
These doubts can be wholly dispelled, but not without a deeper understand-
ing of the grammatical structure (syntax) and meaning (semantics) of argument
forms. This understanding will enable us to see why some arguments are valid and
others not, and it will yield rigorous techniques for settling questions of validity
and related issues.
We begin with some syntactic fundamentals. First, the order of the premises
and minor variations in wording that do not alter meaning are irrelevant to an
argument's form. Thus with regard to the previous argument, for example, we
may omit the 'then', reverse the premises, and adiust the wording a bit without
altering the meaning. The result
the result is no longer modus ponens. Instead, this argument has the form
Q
f
I
thcn P
P
Q
Because the second premise 'P' affirms the consequent of the conditional premise
'If Q, then P', this form is called affirming the consequent. (For parallel reasons,
modus ponens is sometimes called affirming the antecedent.)
Moreover, the argument itself, unlike the two previous arguments, is invalid.
Here is a counterexample: The boss is a scrooge; Clara won't get the raise, period,
whether she quits or not. But she won't quit, because she needs the iob to feed her
kids.
At Ieast one instance of affirming the consequent, then, is invalid. A form
like modus ponens all of whose instances are valid is called a valid form. Any form
that, like affirming the consequent, has at least one invalid instance is called an
invalid form.
But not every instance of an invalid form need be invalid. Here is a valid
instance of affirming the consequent:
28 CHaprrn 2
h ・.
If P, then Q
a
P
Hence, although exchanging letters within the conditional alone changed modus
ponens into affirming the consequent, consistent replacement of letters throughout
a form does not alter the form. The preceding form, for example, results from our
original version of affirming the consequent by replacing 'P' by 'Q' and 'Q' by 'P'
everywhere they occur. Thus it still counts as affirming the consequent.
'We
need to be cautious, however, if we obliterate distinctions between letters.
If in the preceding form, we replace 'P' by 'Q', and make no other changes, we get
If Q, then Q
a
a
This form is still affirming the consequent, but it is also affirming the antecedent.
\We might call it "affirming both the consequent and the
antecedent"!
For brevity, logicians have invented symbols to represent logically significant
terms, that is, logical operators.3 \We shall represent 'if . . . then', for example, by
the symbol '-'. So instead of if P then Q', from now on we may write 'P
To save space, we often write an entire form on one line. In this format we shall
- Q'.
r This is not, of course, an exact definition, but as we shall see later (especially in
Section 9.4), the exact definition of'logical operator' is a matter of dispute.
Crassrcar PRopostttoruaL Loctc: Svttrex 29
use the symbol'l-', often called the turnstile, instead of '.'.', and separate the
premises with commas. Thus modus ponens, for example, is written as 'P - Q,
P l- Q'. An argument form written in this format is called a sequent.
'If . . . then' is not the only logical operator. In this chapter we consider four
others. These correspond roughly to the English expressions 'it is not the case
that', 'and', 'or', and 'if and only if', which we represent respectively by the sym-
bols '-', '&', 'v', and '*'.a (Often 'and' is accompanied by the word 'both' and
'or' by the word 'either'1 these additions are usually for clarity or emphasis and do
not affect the logical meaning.) \What is common to all five operators is that they
apply to propositions to produce more complex propositions. That is why the
form of logic considered in this chapter is known as propositional logic. In later
chapters we will consider operators that apply to entities of other types.
The operators 'if . . . then', 'and', 'or', and 'if and only if' are called binary
or dyadic operators, because they combine two statements into a new statement.
'We
may, for example, use the operator 'and' to combine the separate statements
'it is'Wednesday' and 'it is hot' into a new statement 'it is Wednesday and it is hot'.
The operator 'it is not the case that', by contrast, applies to only one proposition
at a time and hence is monadic or unary. Sre ma5 for example, affix it to the
'Wednesday'
sentence 'it is to form the new sentence 'it is not the case that it is
Wednesday'.
The operator expressed by the symbol '-'is called the negation operator, and
a proposition resulting from its application is called a negative proposition or
negation. The symbol '6C' is the coniunction operator; it combines two proposi-
tions, called coniuncts, into a conjunction. Similarly, 'v' is the disiunction opera-
tor; it combines two propositions, called disiuncts, into a disiunction. The
biconditional operator '*' is the least familiar of the five. It combines two propo-
sitions, which we shall call constituents, into a biconditional. A biconditional
asserts the equivalence of its components in the sense that if either one is true, so
is the other.
These five operators can be combined with sentence letters to produce an
infinity of argument forms. The central aim of formal propositional logic is to
establish methods for deciding which of these forms are valid and which are not.
As a first step toward this goal, we might attempt to evaluate forms infor-
mally, by the method of counterexamples explained in Section 1,.2.That is, given
a sequent, we might by trial and error attempt to produce invalid instances-
instances for which there is a possible situation that makes the premises true but
a
Just as mathematicians use both "' and 'x' to represent multipIication, so
logicians
sometimes use other symbols to represent these operators:
that _l
It is not the case
And
Or (none)
If. . . then l
If and only if =
30 CHeprEn 2
the conclusion untrue. The hope would be that either we find an invalid instance,
thus showing the sequent to be invalid, or we fail to find an invalid instance, but
as a result of our search become familiar enough with the sequent to see that it is
valid.
Consider, for example, the sequent 'P v Q F P * Q'. To test its validity
informally, we consider instances, more or less at random. Suppose we take this
instance:
Exercise 2.1.1
check the following forms for validity informally by attempting ro construct an
instance that has an obvious counterexample. If you can do so, write out the
instance and describe the counrerexample that shows it to be invalid. If not, or if
you see that the argument is valid, simply write 'valid' for that form.
1. P*Q,_PF_Q
Cnssrcll PRopostrtoruel Loctc: Svrurax 31
2.PtP&Q
3. PBcQFP
4. PvQFP
5. PFPvQ
6.P-QrQ-P
7. P- Qf f *--O
8. P-Ql-Q*P
9. P*QFPE Q
10. R-Pr-Q
Exercise 2.1.2
Given that modus ponens is also called affirming the antecedent and modus tollens
is also called denying the consequent, what is the name of the sequent in problem
-1 of Exercise2.l.'1.?
Conditionals also have important variants. They can, for example, U"lr.-
sented in reverse order, provided that the antecedent remains attached to the 'if'.
The statement 'if it rains, it pours', for example, can also be expressed as ,it pours
if it rains'. The form is in each case the same: 'R P'. In either order, the antece-
dent is always the clause prefixed by 'if'.
-
There is one exception. Where 'if is preceded by the term 'only', what it
prefixes is the consequent, not the antecedent. The following four sentences, for
example, all assert the same (true) conditional proposition:
Many people find it difficult to keep the meanings of if'and'only if'distinct. Keep
in mind that 'if' always prefixes antecedents and 'only if' always prefixes conse-
quents, and you should have no trouble.
Given these remarks about 'if' and 'only if', it ought to be clear that the
biconditional operator 'if and only if' may be understood as a conjunction of rwo
conditionals, one expressed by 'if', the other by 'only if'. Hence 'P
- Q, iust means
'(P * Q) ga (Q * P)'. \7e could therefore dispense with the symbol '*' and treat
all biconditionals as conjunctions of two conditionals in this way. But we retain
'e', partly in deference to tradition, partly because it saves writing.
Apart from the optional addition of 'either', the term 'or' has few variants
in English. we may regard it, however, as a component of the important term
'neither...nor'.Etymologicallsthisisacontractionof'noteither...or,;itthus
expresses negated disjunction. The sentence 'it will neither snow nor rain', for
example, may be symbolized as '-(s v R)'. It is also acceptable to symbolize this
statement as '-S & -R', which is logically equivalent to '-(S v R)', though this
symbolization has the disadvantage of failing to reflect the English etymology.
The term 'unless' may be thought of as expressing another two-operator
combination, a conditional with a negated antecedent. 'we will starve unless we
eat' says the same thing as 'if we do not eat, we will starve'. we may thus symbolize
CrasstcaL PRopostrtoltar Loctc: Svrurax 35
Exercise 2.2.1
Formalize each of the sentences below, using the following interpretation scheme:
P the peasants revolt Q the queen hesitates
R
- the revolution will succeed S
- the slaves revolt
- -
1. Either the peasants will revolt or the slaves will revolt.
2. Both the peasants and the slaves will revolt'
3. The peasants and the slaves will not both revolt.
4. If the peasants revolt, then the revolution will not succeed.
5. The peasants revolt if and only if they don't fail to revolt.
6. Only if the peasants revolt will the slaves revolt'
7. The revolution will succeed only if the queen hesitates.
8. If the peasants revolt and the queen hesitates, the revolution will succeed.
9. If the p."srntr revolt, then the revolution will succeed if the queen hesitates.
10. The revolution will not succeed unless the queen hesitates.
11. The peasants will revolt whether or not the queen hesitates.
12. The revolution will succeed if the slaves and the peasants both revolt.
13. If either the peasants or the slaves revolt and the queen hesitates' then the
revolution will succeed.
14. If the peasants revolt but the slaves don't, the revolution will not succeed, and
if both the peasants and the slaves revolt, the revolution will succeed.
15. If the peasanrs revolt if and only if the slaves revolt, then neither will revolt'
Exercise 2.2.2
Use premise and conclusion indicators to determine the premises and conclusions
of the following arguments, then symbolize them in the formal notation of prop-
ositional logic using the sentence letters whose interpretation is specified below.
(The forms of all of these arguments, incidentally, are valid in classical logic.)
SentenceLetter lnterPretation
B Descartes believes that he thinks
E Descartes exists
K1 Descartes knows that he thinks
K2 Descartes knows that he exists
J Descartes is justified in believing he thinks
T Descartes thinks
1. If Descartes thinks, then he exists; for he doesn't both think and not exist'
36 Cuaprrn 2
2. If Descartes thinks, then he exists. Hence he does not think, because he does
not exist.
3. Descartes is justified in believing that he thinks if he knows that he thinks. But
he is not justified in believing that he thinks, so he does nor know that he
thinks.
4. If Descartes knows that he thinks, then he exists. For if he knows that he
thinks, then he thinks; and if he thinks, then he exists.
5. Descartes does not exist. For either he knows that he exists or he doesn't exist;
and he doesn't know that he exists.
6. Descartes believes that he thinks. If he does not think, he does not believe that
he thinks. Therefore Descartes thinks.
7. If Descartes thinks, then he knows that he exists, and if he knows that he
exists, then he exists. Therefore, if Descartes thinks, then he both knows that
he exists and really does exist.
8. If Descartes does not exist, then he doesn't think; so if he thinks, it is not the
case that he does not exist.
9. Descartes neither exists nor does not exist. Therefore Descartes thinks.
10. Descartes knows that he thinks if and only if (1) he believes that he thinks,
(2) he is justified in believing that he thinks, and (3) he does in fact think.
Therefore, if Descartes does not think, then he does not know that he thinks.
The formulas of propositional logic have a grammar, and that grammar (or syn-
tax) may be precisely articulated as formation rules. Formation rules define what
counts as a formula by giving general directions for assembling formulas out of
simple symbols, or characters. They are the rules of grammar for a formal lan-
guage. In order to state the formation rules for propositional logic, we need first
to define the character set for propositional logic-that is, the alphabet and punc-
tuation marks from which the formulas of its language are constructed. We stipu-
late that a character for the language of propositional logic is anything belonging
to one of the following four sets:
letters:
Sentence Capital letters from the English alphabet
Numerals: o"t234s678e
Logicaloperators: - & v + €
Brackets: ( )
The only novelty here is the numerals. These are used to form subscripts for
sentence letters when we want to use the same letter for two different sentences
and need a means to keep the letters distinct. Moreover, without subscripts we
could symbolize no more noncompound sentences than we have capital letters-
that is, twenty-six. And though we are unlikely in practice to need more than
Crnssrcal PRopostttoruaL Loctc: Svrurax 37
twenty-six letters at once, a system of logic should not be subject to such arbitrary
restrictions.
With these ideas in mind, we are ready to state the formation rules-the
rules of grammar for the language of propositional logic; they define the notion of
a grammatical formula by telling how to construct such formulas, starting with
,..rt.n.. letters, and combining them with the operators and brackets.
Then the rule would tell us only how ro generate this one formula '-P'. It would not
tell us how to generate '--P' or '-Q'. If, by contrast, we put the rule this way:
2" If P is a formula' then so is -P
we would be mixing the object language and the metalanguage confusingly' It's not
clear what this means. The Greek says exactly what we want while avoiding these
problems.
38 CHaprrn 2
Notice that the only formation rule that introduces brackets is rule 3. This
means that the only legitimate function of a pair of brackets is to delineate the
scope of some binary operator. In particular, brackets are not used to indicate the
scopes of either sentence letters or the negation operator. Thus, for example, none
of the following expressions count as formulas:
(P) -(P) (-P) -(-P) (All wrong!)
Exerclse 2.3
Sorne of the f0110、 ving expressions are fOrnlulas Of prOpOsitiOna1 10gic.()thers are
not.For thOse that are,cxplain hOw thcy are built up by the forination rules.For
thOse that aren't,explain、 vhy they can't be built up by the fOrnlation rulcs.
1。 (P)∨ (Q)
2。 (P&Q)
3.P&Q
4. ∼∼∼P
5.∼ (Pv(Q&S))
6。 (Φ ―→Ψ )
7.((P&Q)∨ (R&S))
8.P
9。 (P― →P)
10。 (P&Q&R)
cHス PrfR
CLASSiCAL PROPOSIT:ONAL
LOGiC:SEMANTiCS
39
40 CHaprun 3
exists? Suppose he has had a mystical vision in which it seemed to him that God
gave him a message. would that make it true? Suppose that he has had such a
vision and that God really did give him the message. The point of these queries is
to clarify the meaning of the sentence 'James knows that God exists'-or, more
broadly to clarify the meaning of the predicate 'knows' in application to religious
assertions. And the general assumption of the inquiry is that to know the meaning
of a sentence or term is to know which possible situations make that sentence true
or that term truly applicable.
This assumption is often expressed by saying that the meaning of a term is
its truth conditions. The truth conditions for a term are rules that specify the
possible situations in which sentences containing that term are true and the possi-
ble situations in which sentences containing that term are false.
In this section we give a truth-conditional semantics for the five logical op-
erators introduced in Chapter 2.That is, we explain their meanings in terms of the
possible situations in which sentences containing them are true and the possible
situations in which sentences containing them are false.
In doing so, we shall employ the concept of truth ualue. A truth value is
a kind of semantic quantity that characterizes propositions. For now, we as-
sume that there are only two truth values: true, or T, and false, or F. A true
proposition has the value T and a false proposition the value F. Moreover, we
assume that in each possible situation each proposition has one, and only one,
of these truth values. This assumption is called the principle of bivalence. Logics
based on the principle of bivalence and the assumption that meaning is truth
conditions are called classical. The dominant logics in rJ(estern thought have
been classical.
Some philosophers have held that classical logic is universally the best form
of logic, or even the only true logic. This book dissents from that view. In part v
we shall explore reasons for thinking that the principle of bivalence, though appro-
priate for some applications of logic, is less appropriate for others. 'we shall con-
sider truth values other than T and F and the possibility that sentences may have
more than one truth value, or none at all. And in Section 1.6.2, we question even
the idea that meaning has anything to do with truth. There we explore a semantics
that defines the meanings of terms, not as their truth conditions, but as their
assertibility conditions-the conditions under which statemenrs containing these
terms are confirmable by adequate evidence. And beyond that we shall glimpse
still more radical ways of departing from the classical tradition. Each of thise
novel semantic assumptions alters our conception of what valid reasoning is. For
now, however, we present the semantics of ihe logical operators in the ilassical
way, as bivalent truth conditions.
We begin with classical logic for two reasons. First, it is highly established in
the Western logical tradition-our tradition. Second, it is, from a semantic view-
point at least, the simplest logic, and it is best to start with what is simple.
our immediate task, then, is to define the meanings of the five logical oper-
ators in terms of their truth conditions. \We begin with the conjunction operator. If
we conjoin two sentences-say, 'it is wednesday' and 'it is hot'-we obtain a
'wednesday
single sentence ('it is and it is hot') that is true if and only if both
CLassrcal Pnopostrtoruar Loctc: Stveurtcs 41
original sentences were true, and false otherwise. Hence the truth conditions for
conjunction may be stated as follows:
The truth value of a conjunction is T in a given situation iff the truth value
of each of its coniuncts is T in that situation.
and
The truth value of a conjunction is F in a given situation iff one or both of
its conjuncts does not have the value T in that situation.
(The term 'iff is a commonly used abbreviation for'if and only if'.) To understand
these truth conditions is, by the lights of classical logic, to understand what con-
junction means.
All of this is well and good if we aim to state the truth conditions for con-
junction when applied to statements of natural language. But we have taken a step
of abstraction into formal logic. We are no longer concerned primarily with sen-
tences, like 'it's Wednesday and it's hot', but with formulas, like 'W & H'. Simple
sentences have been replaced with sentence letters. But a sentence letter, such as
'W'', has in itself no meaning and is neither true nor false. Of course we can giue it
a meaning by associating it with a particular statement of natural language. But
this we do differently in different contexts. For one problem ''W'' may mean "itt
'Wednesday,"
for another "'Water is HrO." So what can it mean to talk about a
situation that makes a mere formula like ''W 6a H'true?
Two things are needed to make talk about possible situations intelligible in
formal propositional logic. The first is an interpretation of the sentence letters.
Interpretations are given by associating sentence letters with statements of natural
language. The interpretation of a sentence letter may vary from problem to prob-
lem, but within a given problem we keep the interpretation fixed. Thus we may
stipulate, for example, that (for the duration of this example) ''W' means "it's
Wednesday" and 'H' means "it's hot." Let us now, in fact, stipulate this. The
second thing we need to make sense of the notion of a possible situation is the
concept of a valuation:
For a formula, such as ''!7 6c H', that contains two sentence letters, there are
four valuations, as shown in the following table:
W H
T T F F
T F T F
42 Crapren 3
That is, both .!7'and 'H'might be true, ''W' might be true and'H'false,'!f'might
be false and 'H'true, and both 'W' and 'H' might be false. Given an interpretation
of the sentence letters, each valuation defines a situation. For example, given the
interpretation stipulated above, the valuation that assigns T to both ''W'' and'H'
defines a possible situation in which it is both lWednesday and hot. Similarly, the
valuation that assigns T to 'w' and F to 'H' defines a possible situation in which it
is Wednesday but not hot, and so on.
Stipulation of the interpretation, however, though obviously essential for
applying logic to natural language, is inessential from a purely formal point of
view. To state truth conditions for the logical operators, we need only to say how
the truth values of sentences containing them depend on the truth values of their
components, not what the components themselves have been interpreted to mean.
Hence truth conditions for formal logic need concern themselves only with valua-
tions, not with interpretations. A valuation alone is not a possible situation, but
merely a pattern of truth values-the empty form, as it were, of a possible situa-
tion. It has the advantage, however, of being an entity definable with mathematical
precision. If we disregard particular interpretations of sentence letters and think
of abstract valuations rather than possible situations, we enter a realm of formal
thought where everything is sharply defined and clear. The truth conditions for
conjunction now look like this:
The truth value of a conjunction is T on a given valuation iff the truth value
of each of its conjuncts is T on that valuation.
and
Because these more abstract truth conditions are stated in terms of valuations, they
are often referred to as valuation rules.
valuation rules will appear so often from now on that it will be useful to
abbreviate them. we shall use the script letter '"[' to stand for valuations, and, as
in the previous section, Greek capital letters will stand for formulas. Instead of the
cumbersome phrase 'the value assigned to <D by Y', we shall write 'y(O),.Thus, to
say that Y assigns the value T to @, we write simply '"f(O) = T'. Using this nota-
tion, we may state the valuation rules for conjunction more compactly as follows:
The same idea may be expressed in tabular form, listing the four possible combi-
nations of truth value for (D and Y on the left and the resulting truth value for @
6c Y on the right:
OY @&Y
T
T
T
F
T
F
T
F
F
F
F
This is called a truth table. Truth tables are, perhaps, easier to read than valuation
rules. But, unlike the rules, they have the disadvantage of not being generalizable
ro more advanced forms of logic. Rules will prove more useful in the long run,
which is why we emphasize them here.
Let's now examine the truth conditions for the negation operator. If we
attach it to a sentence-Say, 'It's snowing'-u7s gs1 a negated Sentence: for exam-
ple, 'It is not the case that it's snowing'. If the sentence is true, its negation is false.
If the sentence is false, its negation is true. This is vividly apparent when negation
is iterated. The sentence 'It is not the case that it is not the case that it's snowing',
for example, is just an elaborate way of saying 'lt's snowing'; any situation in
which one sentence is true is a situation in which the other is true, and in any
situation in which one sentence is false, the other is false as well. Two negations
"cancel out," producing a statement with the same meaning as the original. By the
same principle, three negations have the same effect as one, four likewise cancel
out, and so on. Negation, then, is simply an operation that inverts truth value.
Hence the truth conditions for negation may be stated precisely as follows:
Y(-o)=Tiffl/(@)*T.
y(-o)=Fiffv(@)=T.
This, according to classical logic, is the meaning of negation.'We can also represent
these rules in a truth table. Since the negation operator is monadic, applying to a
single formula rather than to two, there are only two cases to consider instead of
four: the case in which that formula is true and the case in which it is false. The
table shows that -@ has the value listed in the right column when <D has the value
listed to the left:
valence is thus built into the valuation rules themselves. In fact, throughout this
book I consistently define all semantic ideas in terms of truth and untruth, rather
than truth and falsehood. This saves a good bit of trouble in the metatheoretic
work of Chapter 5 and facilitates a smooth transition to nonclassical logics in
Part V.
Often both valuation rules are stated together in a very compact fashion, as
follows:
y(O & Y) =T iff both y(o) =T andV(Y) =T; otherwise,Y(O & Y) = F
Our formulation says exactly this, but it is more explicit about what "otherwise"
means.
U CHnpren 3
Φ
T
-o
F
F
Let's now consider the truth conditions for 'or'. 'Or' has two meanings. It
canmean"either...or...andpossiblyboth"or"either...or...andnotboth.,,
The first meaning is called inclusive disjunction and the second exclusive disiunc-
tion. This ambiguity is unfortunate. Suppose, for example, that on a true-false
quiz you find the statement
\What should you answer? Four is both even and square. So, you might
argue, it is
not eitber even or square-that is, not just one of these two things- it's both.In
that case you would mark the statement false. on the other hand, you might think
that since four is even (and also since it's square), it's true that it is even or square.
In that case you would mark the statement true.
In neither case would you be wrong, and in neither case would you have
misunderstood anything. But if you marked the statement false, that would mean
you understood the 'or' exclusively, and if you marked it true, that would mean
you understood it inclusively.
This problem might be less acute if we were speakers of Latin. In Latin there
are two words for 'or': 'uel' and'aut'.In most contextsr'uel'more naturally
expresses the idea "either . . . or . . . and possibly
both" (the inclusive sense), and
'aut' tends fo mean "either . . . or . . . and not both" (the exclusive sense of ,or,).
In English we sometimes resolve the ambiguity by using the compound rerm 'and/
or' for the inclusive sense. But both 'or' by itself and 'either . . . or'generally admit
of both readings. when we use them, we or our listeners may not know exactly
what we mean.
This situation would be intolerable in a formal logical language. Formal
logic aims at precision. Its operators must have clear and unambiguous meanings.
Therefore, when we introduce an operator like 'v' we must stipulate precisely what
it means. Logicians usually have found the inclusive sense of 'or' more useful, and
so, by convention, that is the sense they have given to the operator 'v'. In fact, ,v,
is just an abbreviation for'uel'.
Apart from cases in which both disjuncts are true (the cases on which the
inclusive and exclusive senses of 'or' disagree), the truth conditions for 'or' are
clear. If one disjunct is true and the other false (e.g., 'either the sun is a star or the
moon is'), then the whole disjunction is rrue. And if both disjuncts are false (e.g.,
'either the moon is a star or the earth is'), then the disjunction is false. Hence the
valuation rules for the operator 'v' are as follows:
(DY tDvY
T
T
T T
T F F
F T
F
F
The logical operator 'v', then, accurately symbolizes the English 'or' only when
'or' is used in the inclusive sense. In spite of this, we need not introduce a special
symbol for exclusive disjunction, since 'P or Q', where 'or' is intended in the
exclusive sense, may be symbolized in our notation as'(P v Q) Sa -(P 6c Q)'-that
is, "either P or Q, but not both P and Q."
In formalizing arguments involving disjunction, we will for the sake of
simplicity and consistency treat the disjunctions as inclusive, except when
there is strong reason not to. But on those fairly frequent occasions when the
meaning of 'or' is unclear, we should keep in mind that this policy is essentially
arbitrary.
We now turn to the truth conditions for conditional statements. Under what
conditions is Y true? Let's consider the case in which the antecedent <D is false
@
-
(we assume nothing about the consequent Y). Now, though @ is false, the condi-
tional invites us to consider what is the case i/iD, hence to suppose o true' This,
however, yields a contradiction, from which (as we saw in Section 1.3) any prop-
osition validly follows. Take a specific instance: The statement 'Napoleon con-
quered Russia' is false. Given this, it is (in a certain sense) true, for example, that
ii Napoleon conquered Russia, then Caesar conquered the universe. Indeed, if
Napoleon conquered Russia, then anything you like is true-because the fact is
thai Napoleon didn't conquer Russia. Thus it appears that when (D is false, @ -+
Y is true, regardless of the truth value of Y.
Let us next consider the case in which the consequent Y is true. Then,
whether or not O is true, Y is true (trivially). Take a specific instance: The state-
ment 'Iron is a metal' is true. Then any conditional containing 'Iron is a metal' as
its consequent is true. For example, 'If today is Tuesday, then iron is a metal' would
be true-because whether or not it is Tuesday (i.e., regardless of the truth value of
the antecedent), iron is a metal. Thus in general we may infer that (D * Y is true
when Y is true, regardless of the truth value of @.
'We
have now concluded that <D .+ Y is true whenever (D is false or whenever
Y is true. Together these conclusions account for three of the four truth combina-
tions for (D and Y; that is, @ r Y is true whenever O and Y are both true, or is
(D
false and Y is true, or (D and Y are both false. The only remaining case is the one
in which (D is true and Y false. But in this case the conditional is clearly false. If,
for example, it is Tuesday and the weather is not hot, then the conditional 'if it is
Tuesday, then it is hot' is obviously false.
To summarize, we have concluded that @ - Y is true if Y is true (regardless
of the truth value of (D), <D - Y is also true if (D is not true (regardless of the truth
value of Y), and @ .. Y is false if O is true and Y is not true. This covers all
possible cases. Hence the truth conditions for'-' are as follows:
46 Cunprrn 3
OY @-Y
T
T
T
F T
F T
F
F
The conditional defined by these truth conditions is called the material con-
ditional.
If you were unconvinced by the reasoning that led us to the truth conditions
for the material conditional, you are not alone. Many logicians (your aurhor
among them) are troubled by this reasoning.
Consider, once again, the last line on the truth table, the case in which @ and
Y are both false. I argued that the conditional was true in that case, since its
antecedent contradicts the facts, and from a contradiction anything follows. Now
surely conditionals are sometimes true when their antecedents and consequents
are both false. The statemenr
(A) If you are less than an inch tall, then you are less than a foot tall.
for example, is uncontroversially true, though (taking 'you' as referring to you) its
antecedent and consequent are both false. But the antecedent and consequent of
the following ttut.rn.nt are also both false, and yet, unlike statement (A), this
statement seems false:
(B) If you have no lungs, then you can breathe with your eyeballs.
If, as these examples suggest, English conditionals are sometimes true and some-
times false when their antecedents and consequents are both false, then the truth
value of an English conditional must not be determined solely by the trurh values
of its components. Something else must figure into the truth conditions.
An operator which forms compounds whose truth value is strictly a function
of the truth values of the componenrs is said to be truth-functional. The symbols
'6c', '-', 'v', and the material conditional as defined by the truth conditions above
are truth-functional. But we have seen evidence that suggests that 'if . . . then' is
not a truth-functional operator and hence is not the material conditional.
Intuitiveln what makes statement (A) true is not that its antecedent contra-
dicts the facts, but that it is necessary, given that you are less than an inch tall, that
you are also less than a foot tall. correspondingly, what makes statement (B) false
seems to be the lack of just such a necessary connection: It is not necessary given
that you have no lungs, that you can breathe with your eyeballs. This suggests that
an English statement of the form 'if P then Q' is true if and only if such a necessary
connection exists, regardless of the truth values of the components.
CrassrcaLPRoposrrror.rlLLoctc: Sevarurtcs 47
The truth conditions for the material conditional take into account only the
truth values of the antecedent and consequent, not the presence or lack of such a
necessary connection. This leads to anomalies not only in the case in which the
antecedent and consequent are both false, but also in the case in which the ante-
cedent is false and the consequent true. In that case a material conditional is true.
But consider this statement:
If there are no people, then people exist.
Once again, contrary to the truth conditions for the material conditional, this
seems false, and once again,the necessary connection is lacking. This case, in fact,
is especially anomalous, since here the antecedent does not merely fail to necessi-
tate the consequent-it actually necessitates the negation of the consequent.
Further anomalies occur in the case in which the antecedent and consequent
are both true. Consider this example:
If the Mississippi contains more than a thimbleful of water, then it is the
greatest river in North America.
conditionals. The material conditional is therefore not just another way of writing
the English'if . . . then'.2
But then if our aim is to evaluate arguments, which we normally formulate
in English, why bother with the material conditional?
Part of the answer is historical. Beginning with the work of the Scottish
philosopher David Hume (1711-1776), many thinkers, among them some of the
founders of contemporary logic, have doubted the intelligibility of this idea of
necessary connection. As a result, many have found the material conditional
(whose truth conditions, though odd, are at least exact) preferable to English
conditionals, whose truth conditions seem bound up with the suspect notion of
necessity. Indeed, logicians have long dreamed of an ideal language, free of all
ambiguitg unclarity, and dubious metaphysics; and the replacement of the English
conditional by the material conditional offered hope of progress toward that goal.
Early in this centurS Bertrand Russell, Ludwig Wittgenstein, and other prominent
philosophers held that with the creation of such a language the perennial philo-
sophical problems, which they regarded as linguistic confusions, would simply
dissolve. There is indeed much to be said for the replacement of murkier notions
by clearer ones, but in the end we may be left wondering whether we have really
solved the problems or merely changed the subject.
In any case, these early thinkers had little choice but to embrace the material
conditional. They needed some sort of conditional operator, and it was not until
midcentury that logicians began to formulate rigorous and illuminating truth con-
ditions involving ideas of necessary connection. Moreover, the material condi-
tional does mimic English conditionals fairly well in many cases. Like English
conditionals, it is always false when its antecedent is true and its consequent false;
and in the other cases its truth value sometimes agrees and sometimes disagrees
with that of English conditionals. certainly, it offers the best approximation to
English conditionals among truth-functional operators. Moreover, its truth con-
ditions are simple and precise. For these reasons, the material conditional has
become the standard conditional of logic and mathematics.
Lately, however, logicians have formulated a variety of alternative truth con-
ditions that seem to reflect more adequately the meanings of English conditionals.
IUTe shall consider some of these in later chapters.
Unfortunately, none of these
alternatives has won universal acclaim as the true meaning of if . . . then'. That is
why we still bother with the material conditional.
This having been said, however, it must be admitted that the common text-
book practice of symboli zing'if . . . then' in English as the material conditional (a
practice in which this textbook too has indulged) is not wholly defensible. The
material conditional is at best a rough approximation to 'if . . . then', and some
patterns of reasoning valid for the one are not valid for the other. We shall be more
careful about the difference between the material conditional and English condi-
'When
tionals for the remainder of this chapter. considering instances of argument
forms containing the material conditional, we shall not translate the material con-
ditional back into English as 'if . . . then', but instead retain the symbol '-' as a
reminder that its meaning lies solely in its truth conditions, not in what we nor-
mally mean by 'if . . . then'.
It remains to discuss the truth conditions for the biconditional operator'*',
which we have associated with the English expression 'if and only if'. Since, as we
saw in Section 2.2,'if' prefixes antecedents and 'only if' prefixes consequents, the
statement form
oifY
may be symbolized as Y + @, and the form
<D only if Y
as O
- Y. Thus the biconditional, as its name implies, can be understood as a pair
of conditionals-more precisel5 as a coniunction of two conditionals. As a con-
junction, it is true if both conditionals are true, and it is false if either or both are
untrue. Now if @ and Y are either both true or both untrue, both conditionals are
true (by the valuation rules for '*'). But if O is true and Y untrue, then O - Y is
untrue; and if (D is untrue and Y true, then Y * O is untrue. In either of these
cases, the biconditional is false. Thus the truth conditions for the biconditional are
as follows:
qy(O * Y) = T iff eitherV(@) = T and"f(Y) = T, or Y(<D) * T and V(Y) *T.
'y(o * Y) = F iff eitherV(@) = T and V(Y) *T, or V(tD) + T andl/(Y) = T.
These rules yield the following truth table:
OY Φ→ Ψ
T
T
T
F
T
F
F
F
F
The biconditional, in other words, is true if the two constituents have the same
truth value and false if they differ in truth value.
Because the truth conditions for '*' are just those for a conjunction of two
material conditionals, '*' is often called the material biconditional operator, and
where (D * Y is true, <D and Y are called material equivalents. Two formulas, then,
are materially equivalent on a valuation if and only if they have the same truth
value on that valuation.
The material biconditional shares with the material conditional the oddity
of ignoring necessary connections between its components. If its components are
either both true or both untrue a material biconditional is true, regardless of the
existence or lack of existence of necessary or relevant connections between the
50 CHnpreR 3
TABLE 3.1
Valuation Rules for PropositiOnal Logic
The first statement is true because both of its componenrs are true, the second
because both of its components are false. Obviousln then, '*' differs from .if and
only if', just as '-' differs from 'if . . . then'. The material biconditional and
English biconditionals do agree, however, when one component is true and the
other false; here the biconditional itself is surely false.
To summarize: The meanings of the five truth-functional operators of prop-
ositional logic are given by the rules in Table 3.1. These rules constitute the com-
plete semantics for classical propositional logic. For each operator there is a rule
telling when formulas of which it is the main operator are true and a rule telling
when those formulas are false. Together this pair of rules implies each such for-
mula is false if and only if it is not true. Hence collectively the valuation rules
embody the principle of bivalence-the principle that each formula is either true
or false, but not both, on all valuations.
The rules are numbered 1-5. I7e will use this numbering for future reference.
Exercise 3.1
The five operators discussed in this section are a somewhat arbitrary selection from
among many possible truth-functional operators, some of them expressible by
common words or phrases of English. Invent symbols and formulate truth tables
and a valuation rule for binary operarors expressible by these English terms (you
may find it easier to do the truth tables first):
Crassrcal PRopostrtonar Loctc: SrvaNrtcs 51
1. exclusive'or'
2.'...unless...'
3.'neither...nor.
4.'notboth...and '; this is sometimes called the nand operator.
The valuation rules tell us what the operators of propositional logic mean. But
they do more than that. They also enable us to understand why in some cases one
formula must be true if others are true; that is, they enable us to understand why
some argument forms are yalid-and not merely to understand, but to confirm
our understanding by calculation. Because propositional logic makes such calcu-
lations possible, it is sometimes called the propositional calculus.
In Chapter L we said that an argument is valid iff it has no counterexam-
ple-that is, iff there is no possible situation in which its premises are true but
its conclusion is untrue. In this chapter we have shifted our attention from speci-
fic arguments to argument forms. For forms, too, we may define a notion of
counterexample:
And in the actual situation the premise is true but the conclusion isn't.
Conversely, given a counterexample to an instance of a sequent, we can
always construct a counterexample to the sequent by assigning to its sentence
letters the truth values of the corresponding sentences in the counterexample to
the instance. consider, for example, this argument, which is an instance of the
sequent'PFP&Q',:
Bill is a prince.
Bill is a prince & Jill is a queen.
Here is a counterexample to this argument:
Bill is a prince, but Jill, the millert daughter, is a poor but honest maiden,
not a queen.
This counterexample makes 'Bill is a prince' true and 'Jill is a queen' false. we can
turn it into a counterexample to the sequent by ignoring our interpretation of the
sentence letters and assigning these truth values directly to the corresponding sen-
tence letters themselves. The result is the valuation 7 such that T(,p,) = T and
y('Q') = E, which is a counterexample to the sequent.
In this way we can always convert a counterexample to an instance of a
sequent into a counterexample to the sequent itself, and vice versa. This realization
leads us to a new understanding of the concept of a valid argument form. In
section 2.L,we defined a valid argument form as a form all of whose instances are
valid arguments. Thus an argument form is valid iff none of its instances have
counterexamples. But we have just seen that for each possible situation that is a
counterexample to an instance there is a valuation that is a counterexample to the
form, and vice versa. Therefore, to say that no instances ofthe form have counter-
examples is equivalent to saying that the form itself has no counterexamples. Thus
we may equally well de6ne validity for an argument form as follows:
Likewise, since an invalid form is just one that has an instance with a coun-
terexample, and since there is a counterexample to some instance iff the form has
a counterexample, we may likewise redefine the concept of invalidity for an argu-
ment form:
we shall rely on these new definitions from now on. central to both is the
concept used by the valuation rules to define truth conditions: the concept of a
Crassrcar PRoposrrronar Locrc: Srvarurrcs 53
valuation. Thus these definitions illuminate the relationship between the concepts
of validity and invalidity and the valuation rules. The remainder of this chapter
shows how to utilize this relationship to develop computational tests for validity
and other semantic properties.
As a first step in this direction we note that, given a valuation, the valuation
rules enable us to calculate the truth value of a formula or set of formulas from
the truth values assigned to their component sentence letters. For example, given
the valuation V such that V('P') = T, y('Q') = T, and "l/('R') - R we can calculate
the truth value of the formula 'P * (Q & -R)' as follows. Since V('R') + I by the
valuation rule for negation, y('-R') = T. And since both y('Q') = T and y('-R') =
I by the valuation rule for conjunction y('Q & -R') = T. And, finally, since both
'f('P') = T and 1/('Q & -R') = T, by the valuation rule for the material conditional,
"l/('P-(Q&-R)')=t
'We
may list the results of such calculations for all the valuations of a formula
or set of formulas on a truth table. If we do this for the set of formulas that
comprises a sequent, the table will display all the possible valuations of the prem-
ises and conclusion, each as a single horizontal line. We can then, simply by scan-
ning down the table, check to see if there is a line (i.e., a valuation) on which the
premises are true and the conclusion is false. If so, that line represents a counter-
example to the sequent and the sequent is invalid. If not, then (since all the valua-
tions of the sequent are displayed on the table) there is no counterexample and the
sequent is valid. Here at last is a simple, mathematically rigorous test for validity,
one that relies on neither intuition nor imagination!
Let's try it out. Our example will be a sequent expressing modus ponens,
where '-' is now explicitly understood as the material conditional. This sequent
contains only two sentence letters, so it has four possible valuations (both 'P' and
'Q' true, 'P' true and 'Q' false, 'P' false and 'Q' true, both 'P' and 'Q' false), which
we list in the two leftmost columns of the table. Then, beneath each formula of
the sequent, we write its truth value on each of those valuations, like this:
PQ P_Q, P FQ
T
T
T
T
T
F T F
F T T
F T F
F F
F F
Each horizontal line represents a single valuation. For example, the second line
from the bottom represents the valuation Y such that "l/('P') = F and "l/('Q') = T.
To the right, below each formula of the sequent, is listed the truth value of that
formula on 7. On this valuation, for example, 'P - Q' is true. Since the truth table
is a complete list of valuations, if there is a valuation on which the premises are
true and the conclusion is not, it will show up as a line on the table. In this case,
there is no such line, that is, no counterexample. (The only valuation on which
both premises are true is the first one listed, the valuation Y such that V('P') = T
and 1/('Q') = T. But on this valuation the form's conclusion 'Q' is true.) Thus
modus ponens is valid for the material conditional.
54 CHaptn 3
Q Q′ Q
T
T
T
T
T
F
F T
F
F
F
F
On the third valuation listed in the table the premises 'P * Q' and 'Q' are both
true, but the conclusion 'P' is false. This valuation is therefore a counterexample
to the sequent, proving it invalid.
\7e can use the counterexample displayed in the truth table to construct
instances of the sequent that are invalid arguments. Since the valuation which
makes 'P' false and 'Q' true provides a counterexample, we need merely substitute
any sentence that is actually false for'P' and any sentence that is actually true for
'Q' to obtain an invalid instance. Since these are the truth values these sentences
have in the actual situation, a description of the actual situation constitutes a
counterexample to that instance. Let 'P', for example, be the false sentence 'Logic
is a kind of biology' and 'Q' the true sentence 'Logic is an intellectual discipline'.
Then we obtain this instance:
The premises of this argument are true and its conclusion false in the actual
situation.
This technique sometimes yields puzzling instances whose conditional prem-
ise, though actually true, seems false if we confuse '-' with 'if . . . then'. But
keeping the truth conditions for '*' distinctly in mind resolves the puzzlement.
To obtain the values listed under the formulas in the preceding tables, we
just copied them from one of the leftmost columns (if the formula was a sentence
letter) or derived them directly from the valuation rules (in the case of the condi-
tional formulas). With more complex formulas, however, we may need to apply
the valuation rules successively, a step at a time, to calculate truth values for whole
formulas.
Consider the sequent '(P & Q) v (-P & -Q) f -p * -Q'. (To give this some
intuitive content, we might interpret'P' as the statement'The princess dines'and
'Q' as the statement 'The queen dines'. This makes the argument: Either the prin-
cess and queen both dine or neither dines; therefore the princess does not dine if
and only if the queen does not dine.) Ife begin the table for this sequent as before
by listing the four possible valuations of the two sentence letters 'P' and 'Q' in the
two leftmost columns. Now we use the valuation rules to calculate truth values,
starting with the sentence letters and working our way up to more and more
Cmssrcal PnoposrrronaL Locrc: Srvanrrcs 55
complex formulas. The first step is to copy the 'P' column at the left of the table
under each occurrence of the sentence letter 'P' in the formulas and the 'Q'column
under each occurrence of 'Q'. !7here 'P' or 'Q' are directly preceded by '-', how-
ever, we know that their truth values will be reversed, so in these cases we reverse
each truth value in the column we copy. The table now lists the truth values for all
sentence letters or negated sentence letters in the formulas:
Q & Q) (-P&-a)F-P ∼Q
T
F
T
F
T
T
F
F
F
F T
T
T
F
F
T
F
T
Notice that we have written the columns for negated sentence letters under the
negation signs. In general, whenever we are listing the truth values for a complex
formula, we write them under the operator whose scope is that formula.
The next step is to use the valuation rules to calculate the truth values for the
formulas directly joining those whose truth values we have already identified. In
the case of the premise, these are the two conjunctions 'P & Q' and '-P & -Q'.
Remembering that a conjunction is true iff both conjuncts are true, we place a 'T'
beneath '&' on lines where the sentence letters it joins are both true and an 'F' in
all other cases. In the case of the conclusion, the operator joining the'-P' and'-Q'
is the biconditional, and its scope is the entire conclusion. This biconditional is
true on those lines in which '-P' and '-Q' have the same truth value and false
where they differ in truth value. 'We write these values beneath the symbol '*'.
The truth table now looks like this:
T T
T F
T F T
T
F F T
F F F
F T
F F T
F T
T
F T
F F
F
F
F
One step in our calculation remains. The premise is a disjunction of the two
conjunctions whose truth values we have just determined. A disjunction is true if
and only if one or both of its disjuncts are true; otherwise, it is false. Using this
rule, we write the appropriate truth values beneath the symbol 'v':
a (P&a)v(-P&-a)F-P ∼Q
、
T
T
T
F
T
T
F
T
T
F F
F F
F T
T T
F T
F T
T T
F T
F
F
T
´
We also circle the column under the main operator of each formula. Only the
56 CHaprrn 3
circled values, the ones listing the truth values for whole formulas, matter. The
other columns of truth values on the table are merely part of the calculation by
which the circled values were obtained.
As before we read the table by scanning down the columns of truth values
for whole formulas (the circled columns), looking for a valuation on which the
premise but not the conclusion is true. There isn't any; that is, there is no counter-
example. So the sequent is valid.
When a sequent contains two sentence letters, there are four valuations of
that sequent and hence four lines on the truth table. But not all argument forms
contain two sentence letters. Some contain only one, some three or more.'SThere
the number of sentence letters in a sequent is z, the number of valuations of the
sequent) and the number of lines on its truth table, is 2'. Thus a sequent containing
only one sentence letter has 2t = 2 valuations, a sequent containing three has 23 =
8 valuations, a sequent containing four has 2a = 16 valuations, and so on.
It is useful to list the valuations in a standard order. Our convention is as
follows: List all the sentence letters of the sequent horizontally in the top left corner
of the table in alphabetical order. (Where the letters have superscripts the order
remains alphabetical, but letters with lower superscripts are written before those
with higher superscripts, and letters with no superscripts come first of all.)
Then list the valuations under these letters as follows. Beneath the rightmost
letter, write a column of alternating 'T's and 'F's, beginning with 'T', continuing
downward until you have written 2" 'T's and 'F's, where n is the total number of
sentence letters. Then under the next rightmost letter, write another column of 'T's
and 'F's, beginning with 'T', but doubling the number alternated. In other words,
write two 'T's, two 'F's, and so on, downward until again you have written 2"'T's
and 'F's. Now moving to the next rightmost letter (if one remains) and beginning,
as always, with 'T' (because, after all, truth deserves priority over falsehood), write
another column of 2"T's and F's, doubling the alternation again (four'T's, four
'F's, and so on). Keep moving to the left, doubling the number of the alternation
between T's and F's until each letter has a column of 'T's and 'F's beneath it. For
the three letters 'P', 'Q', and 'R', for example, the listing of valuations looks
Iike this:
Rい
Q
T
T T
T
T
F
T
T
F
T
F
T
T
F F F
T F
F T
F
F
Argument forms containing many sentence letters are tested for validity in
just the same way as argument forms containing only one or two. Consider, for
Crassrcal PnoposrrroruaL Locrc: Se varurrcs 57
example, the sequent'(P & Q) - R F (P* R) & (Q- R)'. Once again, we simply
recopy the columns for the sentence letters from the columns to the left, then
calculate the values for formulas of successively larger scope by using the valuation
rules, and finally circle the columns under the main operators. Here is the result:
hK
Rい
nr
pい
Q & Q) t- & (Q R)
T
T
T
T
T T T
T T T
T T T
F T F
F T F T
T F F T
F T T
F T
F T
F T T
F T F
F T F T
T
F F T
F F T
F
F T
F T
T
F
T
F
F
F
F
F
T
T
F
F
F
F
F
F
F
F
The table shows that two valuations are counterexamples to this sequent: the
valuation on which 'P' is true and 'Q' and 'R' are false, and the valuation on which
'P' is false, 'Q'true, and 'R' false. Thus the sequent is invalid.
The significance of these counterexamples can be made clearer by consider-
ing a specific interpretation. Let'P' stand for'The match is lighted', 'Q' for'You
drop the match into a can of gasoline', and 'R' for 'An explosion occurs'. Then the
first counterexample represents a situation in which the match is lighted but you
don't drop it into a can of gasoline so that no explosion occurs, and the second
counterexample represents a situation in which the match is not lighted and you
do drop it into a can of gasoline but once again no explosion occurs.
Let's now consider the truth table for the sequent 'Q i- P * P'. Like some of
the examples discussed in Section 1.3, this sequent lacks relevance; yet it is clearly
valid, as its truth table shows:
p・
a AFP
T
T
T
T
T
T T
F
F F
F F
F
The sequent is valid because there is no valuation on which the premise is true but
the conclusion is not. This, however, is due to a peculiarity of the conclusion: It is
true on all valuationsl it cannot in any way be false. In Section 1.3 we noted that a
logical truth, that is, a statement five in all possible situations, validly follows
from any set of premises. 'P - P' is, of course, a symbolic formula, not a statement.
But if we supply it with an interpretation by assigning to 'P' some specific state-
ment (any statement will do), then 'P * P' expresses a logical truth. In a sense,
then, 'P - P' itself, though a formula and not a statement, is logically true. A
formula that is logically true in this sense is said to be valid:
58 Craprrn 3
column under its main operator contains only 'F's, and a set of formulas is incon-
sistent if and only if there is no horizontal line on which all formulas show a 'T'
beneath their main operators.
In Section 1.3 we observed that any conclusion follows from an inconsistent
set of premises. The medievals called this principle ex falso quodlibet-"from a
falsehood, anything you please." (Actually, this is misleading; the phrase better fits
the relation between the antecedent and consequent of the material conditional; a
false antecedent materially implies whatever you please. But a merely false premise
or premise set does not validly imply all conclusions. From the premise'The Earth
'We
is flat', for example, we cannot validly deduce whatever we please. can, how-
ever, do so from an inconsistent premise or premise set.) Ex falso quodlibet may
be plainly depicted on a truth table. The premises of the sequent 'R -P F Q', for
example, constitute an inconsistent set:
a R ∼P 卜 Q
T F
T T
T F
T T
F F
T
T
T
F
F
F
F
By scanning down the table, we see that there is no horizontal line (valuation) on
which both premises are true so that the set consisting of the two premises is
inconsistent. ObviouslS then, there is no valuation on which both premises are
true while the conclusion is not true. So the sequent is valid.
Corresponding to the two notions of inconsistency defined above are the
following two notions of consistency:
The truth table of a consistent formula has at least one 'T' in the column be-
neath its main operator. The truth table of a consistent set of formulas contains
at least one horizontal line on which there is a 'T' beneath the main operator
of every formula of the set. Some authors use the term 'satisfiable' instead of
'consistent'.
Formulas which are consistent but not valid are said to be contingent:
TABLE 3.2
Semantic Classification of Formulas
TABLE 3.3
Semantic Classification of Sets of Formulas
All formulas fall into one of the three categories: valid, contingent, or incon-
sistent. These are summarized in Table 3.2. A consistent formula, of course, is just
one that is not inconsistent. It is therefore either valid or contingent. We could
also, therefore, have divided all formulas into the two classifications, "consistent"
and "inconsistent," instead of into the three categories described in Table 3.2.
For sels of formulas this twofold classification is generally the most useful. See
Table 3.3.
Truth tables are useful for detecting one other semantic relationship that is
of great importance-logical equivalence:
DEFINITION Two formulas are logically equivalent iff they have the same
truth value on every valuation of both.
'!7ith
respect to truth tables, two formulas are logically equivalent if and only if
the columns under their main operators are identical. Consider, for example, the
formulas '-(P v Q)' and '-P & -Q', which as we saw in Section 2.2 are both ways
Crassrcar PRopostrtottat Loctc: Srualrtcs 61
Q V Q) ∼P & ∼Q
F T
F F F
F F T
T T T
T F T
T T F
T F T
T T
F F
F
F
T
T
T
F
F
T
F
F
F
The truth table reveals their equivalence in that the columns under their main
operators (the circled columns) are identical.
Logical equivalence is significant for several reasons. For one thing, logically
equivalent formulas validly imply one another. That is, an inference from either as
premise
-.r.rse
to the other as conclusion is valid. There can be no counterexample be-
since the two formulas have the same truth values on all valuations, there is
no valuation on which the premise but not the conclusion of such an inference
is true.
Further, since in classical logic meaning is truth conditions and since logically
equivalent formulas have identical truth conditions, it follows that logically equiv-
alent formulas have, for the purposes of classical logic, the same meaning. Thus
.-(p v ,-P &
Q), and -Q, are equally adequate symbolizations for 'neither P nor Q'
because from the viewpoint of classical logic they are synonymous. 'Neither . . .
nor', in other words, ii not ambiguous. Rather, these formulas are two ways of
expressing the single meaning that 'neither . . . nor' has in English. The logical
meaning of 'neithei . . . nor' is, in other words, simply the pattern of truth values
the table for each of these formulas displays.
The material conditional formula 'P * Q' is equivalent to both '-P v Q' and
'-(P & -Q)', as the following table shows:
F
T
T
F
T
T
T
T
T
T
T
F
F
F
F
F
F
T
F T
F
T
F F
T
T T
T T
T T
T
F F
F
T
F
F
we would write
-Pv(-QvR)
Likewise, we could eliminate '*' by taking formulas of the form @ * y as
abbreviations for more complex but equivalent formulas, say, those of the form
(o & Y) v (-@ & -Y) (see problem 1 of Exercise 3.z.4).This would leave us with
only three operators:' -' r' k' rand'v'.
we could reduce our vocabulary still further, either by rewriting (D & y as
-(-(D v -Y) or by rewriting (D v Y as -(-(D & -y) (see problems 2 and 3 of Exercise
3.2.4), thus eliminating either '6a' or 'v'. only two of the five operators, then, are
really needed-either '-' and '&' or '-' and 'v'. And either of these pairs can be
further reduced to a single operator, though not to one of the familiar five. This
final reduction requires either the operator 'neither . . . nor', which is written as a
downward pointing arro% 'f', or as the operator ,not both . . . and,, sometimes
called 'nand', and written simply as 'l'. (Neither symbol belongs to the language of
propositional logic as we defined it in Section 2.3; thus to do this reduction we
would have to adopt a new ser of formation rules.) The truth tables for these
operators are as follows:
Φ Ψ (DIYOY ol Ψ
T T
T F
F F
T T
F T
F
F
T
F
Here we consider the reduction of '&' and '-' to 'l', leaving the reduction to 'f' as
an exercise. This reduction depends on the equivalence of -<D to <Dl<D and the
equivalence of tD & Y to -(<DlY), rhar is, (oly)l(oly), which the following rruth
tables illustrate:
Φ -o ΦIΦ
F
T
F
T
F
OY Φ &Ψ (Φ IΨ ) │ (Φ IΨ )
T T
F T
T
F T
F
r
r
F
T
F
T
r■ ﹁
F
T
F
T
r
(In each case we represent two separate formulas on the same table to show their
equivalence.) Thus we can see that any formula has an equivalent whose sole
Crassrclr PRopostrtoruar Loctc: SEuaruttcs 53
operator is 'l'. We can express the conditional 'P - Q', for example, first in terms
of '-'and'&':
∼(P&∼ Q)
Then the ncgation sign prenxing`Q'may be climinatcd in tcrins of l':
∼(P&(QIQ))
'and`8ご iS arbitrary・
(The Orderin WhiCh We Climinate partiCular OCCurrCnCeS Of`∼ )
∼((PI(QIQ))│(PI(QIQ)))
And inall"We elinlinate the initial negation:
((PI(QIQ))│(PI(QIQ)))│((PI(QIQ))│(PI(QIQ)))
The material conditional! In principle, we could by such means express every
formula of propositional logic solely in terms of the operator 'l'. The unreadability
of the resuli explains why in practice we don't. Thus we will stick with the tradi-
tional five op.rrto.r, chiefly because, though redundant, they give us a reasonably
comprehensible language.
'We can say much with
Yet here we have learned something remarkable:
paltry means if we are willing to tolerate long formulas. Humans generally are
not; but, as we shall see in chapter 10, computers have a different opinion.
EЖ ercise 3.2.1
usc truth tablcs to tcst the fo1lowing argument forins for validity.Write cither
`valid'or`invalid'beside the table to indicate the answe■
1.P∨ Q,P卜 Q
2.P∨ Q,∼ P卜 Q
3。 P→ Q卜 ∼ Q→ ∼P
4.P→ Q卜 Q→ P
5。 P一 Q,∼ P卜 ∼ Q
6。 P→ Q,∼ Q卜 ∼ P
7.P卜 Q→ P
8.∼ Q卜 Q→ P
9。 P卜 P→ Q
10。 ∼ (P・
→ Q)卜 P&∼ Q
ll.P∨ Q卜 P→ Q
12.P&Q卜 P→ Q
13.P8cQ卜 Q8(P
14.P∨ Q,P→ R,Q→ R卜 R
15。 P∨ Q,Q∨ R卜 P∨ R
16.P→ Q,Q一 P卜 P→ Q
64 CHapren 3
t7.Pt-(P&-P)
18. -(P&Q)F-P&-Q
19. -(P v -P) l- Q
20. Pl-(P*Q)*(P&Q)
Exercise 3.2.2
Use truth tables to determine whether the following formulas are valid, contingent,
or inconsistent. ntrrite your answer beside the table.
1
P*-P
2
P*--P
3
P*--P
4
P*-P
5
P 6. --P
7
PvP
8
(P&(P-Q))-Q
9
(P,,
Q)* (Q v P)
0
(P--Q)-(-PvQ)
Exercise 3.2.3
Use truth tables to determine whether the following sets of formulas are consisrent
or inconsistent. !7rite your answer beside the table.
QQtコ
,Q,
P
1
+-p
一 ↑ V oO
Q と Q
P
2
e-p
P
3
-a
P
4
5
Exercise 3.2.4
1. use a truth table to verify that 'P * Q' is logically equivalent both to ,(p
& (Q -* P)'and to'(P & Q) (-p & -e)'. - e)
"
2. Use a truth table to verify that'P & Q' is logically equivalent to .-(-p v -e),.
3. Use a rruth table to verify that 'p v e' is logically equivarenr to '-(-p sa -Q),.
4. Use a truth table to verify that 'Q r P' and '-Q p,, which are both wayi of
symbolizing'P unless Q', are equivalent.
-
5. Find equivalents for the forms -<D and @ & y in terms of 'f', and show that
they are logical equivalents by constructing the appropriate truth tables.
6. Find a logical equivalent for (D v Y in terms of 'l', and demonstrate the equiva-
lence with a truth table. Do rhe same thing in rerms of ,f,.
Cnssrcal PRopostttorual Loctc: Se verurtcs 55
A semantic rree is a device for displaying all the valuations on which the formula
or set of formulas is true. Since classical logic is bivalent, the valuations on which
the formula or set of formulas is false are then simply those not displayed. Thus
trees do the same job as truth tables. But they do it more efficiently; especially for
long problems, a tree generally requires less computation and writing than the
corresponding truth table. A truth table for a formula or sequent containing z
sentence letters has 2" lines. For n = 10, for exampler 2" = 1'024-a good many
more lines than we are likely to want to write. But a tree for a formula or sequent
with ten sentence letters (or even more) may fit easily within a page. Moreover, as
we shall see in Section 7.4, trees have the advantage of being straightforwardly
generalizable to predicate logic, which truth tables are not.
Suppose, for instance, that we want a list of the valuations on which the
formula '-P 6C (Q v R)' is true. To obtain this by the tree method, we write the
formula and then begin to break it down into those smaller formulas which, ac-
cording to the valuation rules, must be true in order to make '-P & (Q " R)'true.
Now ,-P & (Q v R)' is a coniunction, and a conjunction is true iff both of its
conjuncts are true. So we write '-P & (Q v R)', then check it off (to indicate that
it has been analyzed), and write its two con,uncts beneath it, like this:
/ -P&(QvR)
-P
Q'R
A formula which has been checked off is in effect eliminated. We need pay no
further attention to it. I07hat remains, then, are the two formulas '-P' and'Q t R'.
'-P' is true on just those valuations on which'P' is false. But we still need to analyze
'Q v R'.
Now, whereas a coniunction can be true in only one way (both conjuncts are
true), there are two ways in which a disjunction can be true: Either the first disjunct
is true or rhe second disjunct is true (or both-but this possibility is in effect
already included in the other two, as will be explained shortly). Hence to analyze
.Q v R', we check it and split our list into two branches, the first representing
valuations in which 'Q' is true, the second representing valuations on which 'R' is
true) as follows:
/ -P6.(Q"R)
-P
/ QvR)
,,4'.
,/\
aR
create by this
It is because lists may "branch" in this way that the structures we
trees. (But these trees grow downward!) ltr7hen all
procedure are called semantic
65 CHaprrR 3
formulas other than sentence letters or negated sentence letters have been checked
off, as they have here, the tree is finished. This tree contains two "branches," or
patbs, one running from'-P & (Q v R)'to'Q', the other from'-P & (e v R)'
to'R'.
Now we scan along each path, looking for sentence letters or negated sen-
tence letters. Along the first path we find'-P' and 'Q'. This shows that ,-p Ac (e
R)' is true on those valuations which make both '-P' and 'Q' true, that is, those "
valuations which make 'P' false and 'Q' true. But 'R' does not appear either by
itself or negated along the 6rsr path. This indicates that if 'p' is false and 'e' true,
'-P & (Q v R)' is true, regardless of whether 'R' is true or false. The first path,
then, represents these two valuations:
pい
Q
F
T
T
F
Checking the second path for sentence letters or negated sentence letters, we find
'-P' and 'R'. This means that '-P & (Q v R)' is also true on valuations in which
'P' is false and 'R' is true, regardless of the truth value of 'e', which does not
appear, either alone or negated, along that path. Hence the second path represents
these valuations:
Rい
Q
F
T
F
Notice that there is some redundancy here. Both paths represent the valuation on
which 'P' is false and both 'Q' and 'R' are true. This is what I meant when I said a
few paragraphs back that the possibility of both disjuncts being true is in effect
already included in the possibilities of either disjunct being true. Thus together the
two paths represent three valuations, not four:
Q R
F
T
F
F
F
These are precisely the lines of the truth table on which '-P & (Q v R)' is true;
they represent all the valuations which make this formula true. In this way the tree
procedure accomplishes exactly what truth tables do.
Crasstcar PRoposrttoNar Loctc: SEvarurtcs 67
Now '-(P & Q)' is true iff 'P & Q' is false. By valuation rule 2, 'P & Q' is false iff
either'P'or'Q'or both are untrue-that is, by valuation rule 1, iff either'-P'or
'-Q'o. both are true. Thus we may check'-(P & Q)' and split our list into two
branches, the first representing valuations on which '-P' is true, the second repre-
senting valuations on which '-Q' is true:
/ -(P&Q)&p
-(P & Q)
P
,/\
,/\
,/\
-P
x
-a
!7e have placed an 'X' at the bottom of the left branch because the path it
(-p'-6ontains both'P'
represents-the path extending from '-(P Sc Q) ec p' 1e
and '-P' and must be closed. This path represents no valuations. The path that
follows the right branch, however, does not close. On it we find 'P' and '-Q'. Since
these are the only letters in our initial formula, '-(P & Q) & P', that formula
is true on only one valuation, namely, the valuation on which 'P' is true and 'Q'
is false.
Sometimes all paths close. This indicates that the initial formula or set of
formulas is not true on any valuations, that is, is inconsistent. Consider, for ex-
ample, the formulas 'P, Q', '-P', and'-Q', which together form an inconsistent
set. We may set them down in a vertical list and apply the same procedure as
above:
/PVQ
∼P
O、X
P X
58 CHeprrn 3
When we check 'P v Q' and analyze it into its two truth possibilities,'P'and'Q',
we see that each of the resulting paths contains both a formula and its negation:
'P' and '-P' on the path that branches to the left, 'Q' and '-Q' on the path that
branches to the right. So we place an 'X' at the bottom of each path to indicate
that it is closed.
'lfith both paths closed, there is nothing more to be done. The tree
is complete, and it shows that there are no valuations on which the formulas of
our initial list ('P v Q', '-P', and '-Q') are all true.
In summary, then, to construct a semantic tree, list the formula or set of
formulas to be tested in a single column. Then check off a complex formula on the
list, writing at the bottom of the list simpler formulas that would have to be true
if this complex formula were true. If the complex formula can be true in more than
one way, split the list and display formulas representing each of these ways on
a separate "branch." Then repeat this procedure for other complex formulas in
the list. Eventually along each path of the tree one of two things will occur. Either
(1) all the formulas along that path of the tree will be simplified inro sentence
letters or negations of sentence letters or (2) some formula and its negation both
will appear. In the first case, the path displays one or more valuarions on which
the initial formula or set of formulas is true-namely, those valuations on which
isolated sentence letters along that path are assigned the letter T, negated sentence
letters along that path are assigned the value F, and sentence letters not appearing
either negated or unnegated along the path are assigned either T or F. In the second
case, the path is a dead end and represents only a failed attempt to construct a
valuation. We close it with an 'X'.
The following terminology will be helpful:
In order to apply the tree procedure formally, we need to specify exact rules
by which complex formulas are to be analyzed into simpler components. There is
nothing surprising here. The tree rules simply mimic the valuation rules. This is
why trees are called semantic trees. They are simply a perspicuous way of display-
ing the semantics for any formula or (finite) set of formulas.
There are ten tree rules, two for each of the valuation rules (that is, one for
the truth clause and one for the falsity clause of each rule). Each tree rule is listed,
Crassrcal Pnoposrrroruar Locrc: Sevarurtcs 69
along with the clause of the valuation rule to which it corresponds, in Table 3.4.
The trees we have done so far have exemplified the negation, coniunction, negated
conjunction, and disjunction rules.
The conjunction rule, for example, was the one we used to analyze'-(P 6a
Q) & P' into its components:
l. /-(P&Q)&P Given
2. -(P&Q) 1&
3. P 1&
Here we have repeated the first step of the tree for '-(P & Q) 6a P', annotating it
by numbering and labeling the lines-a procedure which, having named the rules,
we will follow from now on. Line 1 is marked 'given'to indicate that it is the given
formula. Lines 2 and 3 are marked '1 &'to show that they are obtained from line
1 by the conjunction rule.
In the second step of this tree, we used the negated conjunction rule to show
that there are two ways in which 'P & Q' can be false-namelS if 'P' is false or if
'Q' is false.
l. /-(P&Q)&P Given
2. -(P&Q) 1&
3. P 1&
4. -P -a 2-U
Finally, we complete the tree by closing the left branch with the negation rule.
From now on we will annotate the 'X' by writing the line numbers on which we
found the formula and its negation, which closed the path-in this case, lines 3
and 4:
Q Q
&
& &
P
n
1 ■ つ を う0
r
一 一 P
Given
V
1&
1&
4. -P -a 2-k
5. X3,4
As noted previouslS this tree shows that '-(P & Q) e< P' is true on only one
valuation-namely, the valuation on which'P'is true and 'Q'is false.
Because a tree for a formula, a set of formulas, or a sequent displays all the
information contained in a truth table for that formula, set of formulas, or se-
quent, any test that can be performed by a truth table can also be performed by a
tree. To test a sequent for validity for example, we must determine whether there
is a valuation that makes its premises but not its conclusion true. This would, of
course, be a valuation on which both the sequent's premises and the negation of
70 CHapreR 3
TABLE 3.4
The Ten Tree Rules
its conclusion are true. We may construct a tree to search for just such valuations
by starting with the argument's premises and the negation of its conclusion.
Let's test the sequent 'P -* (Q & (R v S)) l- P * Q'. Since it contains four
sentence letters, the truth table would take 2a = 16 lines-a laborious task. The
tree method is much more efficient. We list the premise and the negation of the
conclusion, labeling them as such to the right, and then analyze them by mechan-
ically applying the tree rules. Here is the result:
Crassrcar Pnopostrtonar Loctc: Srvarurtcs 71
TABLE 3.4
The Ten Tree Rules(contFnue0
4 ν(Φ ―→Ψ)=T iff either γ(Φ )≠ T Conditional (-*) If an open path contains
or γ(Ψ )=■ Or bOth・ an unchecked formula of the form ((D -
Y), check it and split the bottom of each
open path containing this newly checked
formula into two branches; at the end of
the first write -@, and at the end of the
second write Y.
γ(Φ ―→Ψ)=F iff both γ(Φ )=T Negated Conditional (--) If an open
and γ(Ψ )≠ T path contains an unchecked formula of the
form -(<D - Y), check it and list (D above
-Y at the bottom of every open path that
contains this newly checked formula.
57(Φ ←→Ψ)=T iff eithcr γ(Φ )=T Biconditional (*) If an open Path
and ν(Ψ )=■ Or γ(Φ )≠ T and contains an unchecked formula of the form
γ(Ψ )≠ ■ (O * Y), check it and split the bottom of
each open path containing this newly
checked formula into two branches; at the
end of the first list O above Y, and at the
end of the second list -@ above -Y'
γ(Φ → Ψ)=F iff either γ(Φ )=T Negated Biconditional (-*) If an open
and V(Ψ )≠ ■ Or γ(Φ )≠ T and path contains an unchecked formula of
γ (Ψ )=T the form -(O * Y), check it and split the
bottom of each open path containing this
newly checked formula into two branches;
at the end of the first list <D above -Y, and
at the end of the second list -<D above Y.
4
Premise
/∼ (P→ Q) Negation of conclusion
)--
j ´
P
∼Q
4
/-+
P
5
”X
/Q&(R∨ S) 1-→
6
う0
ξυ
Q 58ζ
7
R∨ S 58こ
8
X4,6
Both paths close, so there is no valuation which makes the premises but not the
conclusion true. Hence the sequent is valid.
72 CHaprrn 3
Notice that we closed the right branch without analyzing 'R v s'. This is
permissible. Any path may be closed as soon as a formula and its negation both
appear on it. Analyzing 'R v s' would have split this path into two new paths, but
each of these still would have contained both 'e' and '-e' and hence each still
would have closed. Closing a path as soon as possible saves work.
It also saves work to apply nonbranching rules first. when I began the tree,
I had the choice of analyzingeither 'P
- (e & (R v S))' or '-(p - e)' first. I chose
the latter, because it is a negated conditional, and the negated conditional rule does
not branch. If I had analyzed 'P
- (Q Bc (R v s))', which is a conditional, first,
then I would have had ro use the conditional rule, which does branch. Then when
I analyzed'-(P * Q)', I would have had to write the results twice, once at the
bottom of each open path. Analyzing 'P * (e & (R v S))'first is not wrong, but it
requires more writing, as can be seen by comparing the resulting tree with the
previous tree:
/(P⇔ Q) Premise
2
/(P― R) 2--
P Q
4
⇔
∼P
5
⇔
P R
/
⇔
0
2 υ う0
∼P P
∼P
ワ′ ハ
∼R R ∼R
ё
X4,6 X4,6
Crassrcar PRopostrtotar Loctc: Srvarurtcs 73
The leftmost and rightmost branches remain open. The leftmost branch reveals
that the premise '(P * Q)' is true and the conclusion '-(P - R)' false (because its
negation is true) on the valuation on which 'P', 'Q', and 'R' are all true. This
vaiuation, in other words, is a counterexample to the sequent. The rightmost
branch reveals that the valuation on which 'P', 'Q', and 'R' are all false is also a
counterexample. Thus the sequent is invalid.
If we begin a treq not with premises and a negated conclusion, but with a
single formula or set of formulas, as we did in the first examples of this section'
the tree tests this formula or set of formulas for consistency. If all paths close, there
is no valuation on which the formula or set of formulas is true, and so that formula
or set is inconsistent. If one or more paths remain open after the tree is finished,
these represent valuations on which the formula or all members of the set are true,
and so the formula or set is consistent.
Trees may also be used to test formula.s for validity. The easiest way to do
this is to search for valuations on which the formula is not true. If no such valua-
tions exist, then the formula is valid. Thus we begin the tree with the formula's
negation. If all paths close, there are no valuations on which its negation is untrue
so that the original formula is true on all valuations. Consider, for example, the
formula '(P -Q) * -(P & -Q)'.'When we negate it and do a tree, all paths close:
/∼ (P→ Q)1
一
2
/(P→ Q)
/∼ (P&∼ Q)1
一
3
/∼ ∼(P&∼ Q)
4
P&∼ Q
3∼ ∼
↓
P 一
つ “ ︵∠
一
5
P 48`
↓
Q
一
6
∼Q 48こ
Q
0、X
3∼ &
r ′ヽ
” X
”
” Ⅸ
∼P
´ヽ
7。 2-→
7′
Therefore, since there is no valuation on which '-((P - Q) * -(P & -Q))' is true,
'(P Q) * -(P & -Q)' is true on all valuations, that is, valid.
- Notice that I closed the rightmost path before '--Q'was fully analyzed. This
is a legitimate use of the negation rule. If the path had not closed, however, the
tree would not be finished until the negated negation rule was applied to '--Q'.
Here is a list of some of the ways in which trees may be used to test for
various semantic properties:
constructing trees is just a matter of following the rules, but there are a few
common errors to avoid. Keep these in mind:
The rules for constructing trees apply only to whole formulas, not to their
parts. Thus, for example, the use of --shown below is not permissible:
l
P
-* --Q Given
V
2
P Q -- (Wrong!)
V
1,
A rule applied to a formula cannot affect patbs not containing that formula.
Consider, for example, the following incomplere tree:
1. ,t P v (Q v R) Given
2. P1v QvRlv
Here the formula 'Q v R' at the end of the right-branching path remains to
be analyzed. The next step is to apply the disjunction rule to this formula.
In doing so, we split this right-branching path but add nothing to the path
at the left, for it does nor contain 'Q ,, R' and is in fact already finished.
The negation rule applies only to formulas on tbe same path.In the follow-
ing tree, for example, both 'P' and '-P' appear, but neither path closes be-
cause the formulas don't appear on the same path:
1. ,u P-P Given
2. ∼Pl― → Pl― →
Cussrcar PRopostrtorunL Loctc: Sevarurtcs 75
Exercise 3.3.1
Redo Exercises 3.2.1, 3.2.2, and 3.2.3 using trees instead of truth tables.
Exercise 3.3.2
1. How might trees be used to prove that two formulas are logically equivalent?
Explain.
2. To prove a formula valid using trees, we construct a tree from its negation. Is
theie a way to prove a formula valid by doing a tree on that formula without
negating it? Explain.
How can this be? The answer lies in the distinction between valuations and possi-
ble situations.
Suppose we let 'Sr' stand for 'Some men are saints' and 'Sr' for 'Some saints
are men'. Then we can represent the form of the argument as 'Sr - Sz, Sz F Sr'.
Here is its truth table:
51 s2 S,*52, 52 FSt
T
T T
T
T T F
T F T
F T
F T
F
F
T
F
F
F
The valuation in which 'Sr' is false and 'Sr' true is a counterexample to the sequent
or argument form. Btt the corresponding situation-the one in which 'Some men
76 CrapruR 3
are saints' is false and 'Some saints are men' is 11us-i5n'1 a counterexample to the
argument because it isn't a possible situation. The very idea of a situation in which
some men are saints but it is not the case that some saints are men (i.e., no saints
are men) is nonsense. of course we can easily find other interpretations of 's,'and
'Sr'-and consequently other instances of this form-to which the valuation that
makes 'S,' false and 'Sr' true provides a genuine counterexample, even an actual
counterexample. But on this particular interpretation, that valuation corresponds
to an lzapossible situation. (Incidentallg so does the valuation that makes'Sr'true
and 'Sr' false.) An lzzpossible situation, if it even makes sense to talk about such a
thing, cannot be a counterexample.
Depending on how we interpret the sentence letters, then, a particular valu-
ation may or may not correspond to a possible situation. For many interpretations,
all valuations correspond to possible situations. For example, if we let 'Sr' stand
for 'It is sunny' and 'Sr' for 'lt is Sunday', every line on the truth table above (every
valuation) represents a possible situation, and the valuation on which 'Sr'is false
and 'Sr' true represents a counterexample to the argument as well as to the form.
In such cases, the statements corresponding to the sentence letters are said to be
logically independent. But where the statements corresponding to the sentence
letters logically imply one another or exclude one another, in various combina-
tions, some valuations represent impossible situations.
Such nonindependent statements as 'Some men are saints' and'Some saints
are men' have interrelated semantic structures that are not represented in propo-
sitional argument forms in which they are symbolized simply as sentence letters.
(In this case, the semantic structures in question are relationships among the logi-
cal meanings of the words 'some', 'men', and 'saints'.) In chapter 5 we shall begin
to formalize the semantic structures of such statements, and we shall redefine the
notion of a valuation so that it reflects more of these semantic structures and yields
a more powerful and precise logic. Later we shall explore ways of creating logics
that are more powerful and precise still. But at no point shall our concept of a
valuation become so sophisticated that a valuation may never represent an impos-
sible situation-which is to say that at no point do we ever achieve a formal
semantics or formal logic that reflects all the logical dependencies inherent in
natural language.
Certain consequences of this disparity between valuations and possible situ-
ations, between formal and informal logic, will haunt us throughout this book:
An invalid sequent may have valid instances. The reason for this we have
already seen. The counterexamples to the sequent may on some interpreta-
tions represent impossible situations so that there are no possible situations
which make the corresponding argument's premises true while its conclu-
sion is untrue. No argument is valid because of havingan invalid form, but
an argument may be valid in spite of having an invalid form, because of
elements of its semantic structure not represented in the form.
A contingent formula may have valid or inconsistent instances. A contin-
gent formula is true on some valuations and false on others. But on some
Cussrcal Pnopostttoruar Loctc: Srvnnrtcs 77
All instances of a valid sequent are valid arguments. A valid sequent has no
valuation on which its premises are true but its conclusion is not true. Some
valuations may on a particular interpretation correspond to impossible situ-
ations. Yet since a valid sequent has no valuations representing situations
(possible or impossible) in which the premises are true and the conclusion is
false, none of its valuations represents a possible situation that is a counter-
example to the instance. Hence, if a sequent is valid, all of its instances must
be valid as well. Valid sequents are, in other words, perfectly reliable pat-
terns of inference.
All instances of a valid formula are logical truths. A valid formula is a for-
mula true on all valuations. Even if on a given interpretation some valua-
tions of such a formula do not represent possible situations, the formula is
still true on all the others and hence true in all the situations that are possi-
ble. Therefore any statement obtained by interpreting a valid formula must
be true in all possible situations. That is, it must be a logical truth.
All instances of an inconsistent formula are inconsistent statements. An in-
consistent formula is true on no valuations. Hence, even if on a given inter-
pretation some of its valuations represent impossible situations, still the
formula is true on none of the remaining valuations which represent possi-
ble situations. Therefore any statement obtained by interpreting an incon-
sistent formula is not true in any possible situation.
Under the same interpretation, logically equivalent formulas have as their
instances logically equivalent statements. Logically equivalent formulas are
formulas whose truth value is the same on all valuations. Once again, even
78 CHeprEn 3
CLASSiCAL PROPOSIT10NAL
LOGIC:INFERENCE
Most people can at best understand arguments that use about three or four prem-
ises at once. For more complicated arguments, we generally break the argument
down into more digestible chunks. Beginning with one or two or three premises'
we draw a subconclusion, which functions as a stopping point on the way to the
main conclusion the argument aims to establish. This subconclusion summarizes
the contribution of these premises to the argument so that they may henceforth be
forgotten. This subconclusion is then combined with a few more premises to draw
, Gther conclusion, and the process is repeated, step by small step, until the final
conclusion emerges. The following example illustrates the utility of breaking com-
plex inferences down into smaller ones:
The meeting must be held on Monday, Wednesday, or Friday.
At least four of these five people must be there: Al, Beth, Carla, Dave,
and Em.
Em can't come on Monday or Wednesday.
Carla and Dave can't both come on Monday or Friday, though either of
them could come alone on those days.
AI can come only on Monday and Friday.
The meeting must be held on FridaY.
79
80 CHeprrn 4
and
carla and Dave can't both come on Monday or Friday, though either of
them could come alone on those days.
we can conclude
combining this with the previously derived conclusion that the meeting can,t be
held on'Wednesday and with the premise
conclusion is written on a new numbered line, with the line numbers of the prem-
ises (either assumptions or previous conclusions) from which it was deduced listed
to the right. Here is our reasoning recorded as a proof:
1. The meeting must be held on MondaS lU7ednesday or Friday. A
At least four of these five people must be there: Al, Beth,
2. Carla,Dave, and Em. A
3. Em can't come on Monday or Wednesday. A
Carla and Dave can't both come on Monday or Frida5 though
4. days.
either of them could come alone on those A
5. AI can come only on Monday and Friday. A
5
6. Neither Al nor Em can come on Wednesday.
6
7. The meeting can't be held on lffednesday.
4
3
8. Em and either Carla or Dave can't come on Monday.
8
9. The meeting can't be held on Monday.
9/
1
7
10. The meeting must be held on Friday.
The series of conclusions is listed on lines 5-10. None of these conclusions is
drawn from more than three premises. Each inference is plainly valid. The proof
ends when the desired conclusion is reached. In the remainder of this chapter, we
explore a more formal version of this proof technique.
Exercise 4.1
Analyze each of the following arguments into simple inferences involving at most
three premises each, and write the analyzed argument as a proof. Each inference
in this proof should be obviously valid in the informal sense of validity discussed
in Chapter 1, but it need not exemplify any prescribed formal rule. (Some simple
formal inference rules are introduced in the next section.) There is not iust one
right answer; each argument may be analyzed in many ways.
1, If the person exists after death, then the person is not a living body.
The person is not a dead body.
Any body is either alive or dead.
The person exists after death.
The person is not a body.
2. r is an odd number.
x+Y=25'
x>3,
30/x is a whole number.
r<10.
! =20'
3. You will graduate this semester.
In order to graduate this semester, you must {ulfill the humanities require-
ment this semester.
You fulfill the humanities requirement when and only when you have taken
and passed either (1 ) two courses in literature and a single course in either
82 CHaprrR 4
P*(Q-(S*T)),RP-Q,S l- T
The first step is to write the assumptions in a numbered list, indicating that they
are assumptions by writing an '/t' to the right of each:
1. P-(Q-(S_T)) A
2.P A
3. p-e A
4.S A
Then we look for familiar inference patterns among the premises. For example,
from premises 2 and 3, we may infer by modus ponens the conclusion Q. So we
write this as a conclusion, listing to the right the line numbers of the premises from
which it was inferred and the form or rule of inference by which it was inferred:
5. Q 2,3 modus ponens
The formula 'P', which is assumed on line 2, is also the antecedent of the condi-
tional assumption on line 1. Though the consequent of this conditional is a com-
plex formula, rather than a single sentence letter, we still recognize here another
instance of modus ponens. So we draw the conclusion:
6. Q - (S
- T) 1, 2 modus ponens
(Ii(e have dropped the unnecessary outer brackets, as usual, and will continue to
do so without comment from now on.) Now lines 5 and 6 can be combined to
obtain yet another conclusion:
CrassrcaL Pnopostrtoruar Loctc: llrrRrruce 83
7. S- T 51 5 modus ponens
This is the conclusion we wanted to establish. And now we have succeeded. For
by showing that it is possible to get from the assumptions of the sequent to its
conclusion by simple steps of valid reasoning, we have shown that the sequent
itself is valid.
To see why it is valid, consider a preliminary conclusion c, validly drawn
from some initial set of premises. Now let new premises be used together with C,
to validly draw a second conclusion Cr. Since C, was validly drawn, by the defini-
tion of validity C, is true on any valuation on which the original premises are true'
And similarly since C, validly follows from C, together with the new premises, C2
is true on any valuation on which both C, and the new premises are true. But since
C, is true on any valuation on which the original premises are true, C, is true on
any valuation on which the original premises and also the new premises are true.
Hence, since C, is true on any valuation on which both C, and the new premises
are true, C, is true on any valuation on which both the initial premises and the
new premises are true. That is, the inference from the initial premises together with
the new premises to C, is valid. Further, if we were to add still more premises and
validly draw yet a third conclusion C., the same reasoning would show that the
inference combining all three sets of premises to the conclusion Ci is also valid.
And so it goes. Thus, by stringing together valid inferences, we prove the validity
of the infeience whose premises are all the assumptions made along the way and
whose conclusion is the final conclusion of the string.
In the rest of this section, we shall show how to break down any valid
sequent in propositional logic into a sequence of simple and (more or less) obvi-
ouily validparterns of reasoning. Such sequences, as exemplified by lines 1-8
IJfe
above, are c;lled proofs. Modus ponens is not the only pattern used in proofs.
shall construct proofs in propositional logic by the so-called natural deduction
method, which utilizes ten distinct patterns of reasoning, or rules of inference (or
inference rules), of which modus ponens is one. (There are many other methods of
proof, which use different types and numbers of rules, though for classical logic,
at least, they all yield the same results. Some of the alternative methods are dis-
cussed in Section 4.5.)
Proofs, of course, are only as credible as their inference rules. As we intro-
duce each rule, we shall verify its validity using the semantics developed in Chapter
3. This will enable us to see that our proof technique is sound-that is, that if we
start with assumptions true on some valuation, we shall always, no matter how
many times we apply these rules, arrive at conclusions that are likewise true on
that valuation. Thus a proof establishes that there are no counterexamples to the
sequent of which it is a proof; it is a third formal method (in addition to truth
tables and trees) for showing that a sequent is valid. In Section 5.10 we shall show
that the entire system of rules introduced here is not only sound but also
complete-that is, capable of providing a proof for every valid sequent of propo-
sitional logic.
84 Cnaprrn 4
Given <D
- Y and @, infer Y.
(D and Y may be any formulas, simple or complex. For example, in the inference
from assumptions 1 and 2 to conclusion 6 in the proof above, @ is .P' and y is
'Q*(S-T)'.
Modus ponens is clearly valid, as we can see by examining its truth table.
That is, no matter what the truth values of @ and Y may be, it can never happen
that @ * Y and (D are true but Y is untrue:
OY o Ψ orY
T
T
T
F
F
T
F
In addition to modus ponens, we shall introduce nine other rules, for a total
of ten-two for each of the five logical operators. For each operator one of the
two rules, called an introduction rule, allows us to reason to (introduces) conclu-
sions in which that operator is the main operator. The second rule allows us to
reason from premises in which that operator is the main operator; it is known as
the operator's elimination rule, because it enables us to break a premise into its
components, thus "eliminating" the operator.
Modus ponens is the elimination rule for the conditional. Given a formula
@, it allows us to "eliminate" the conditional operator from @ * y and wind up
just with Y. In doing proofs, then, we shall call modus ponens conditional elimi-
nation, which we abbreviate as '*E'. officially, we state the rule of modus ponens
as follows:
conjuncts are true on that valuation. Hence there is no valuation on which o & y
is true and either of its conjuncts is untrue.
The following proof for the sequent 'R
&E and *E: - (P & Q), R F Q' exemplifies both
1. R*(P&Q) A
2.R A
3. P&Q 1,2-E
4. Q 3&E
As before, we begin by writing the assumptions on numbered lines (lines l and2)
and marking them with an '.Ff to indicate that they are assumptions. '*E' (modus
ponens) applied to lines 1 and 2 gets us the conclusion 'P & Q' at line 3, and from
this by conjunction elimination we obtain the desired conclusion ,e,.
Lett now consider the conjunction introduction rule. This rule enables us to
infer conclusions whose main operator is a conjunction:
A
P
2
1&E
3
1&E
4
2,3 Ul
Starting with the assumprion 'P & Q', we break it into its components at lines 2
and 3 by &E, then introduce the desired conclusion by &I (whose purpose, re-
' member, is to create conjunctive conclusions) at line 4. Intuitively, the reasoning is
this: Given that the conjunction 'P & Q' is true, 'p' is true and 'e' is true *"11.
But then the conjunction'Q & P' is also true. "r
The order in which the premises are listed is irrelevant to the application of
a rule of inference. Thus, even though 'P' is listed on line 2 of this proof and 'e'
on line 3, we may legitimately infer 'Q & P', in which 'Q' comes firsr.
Moreover, the use of two different Greek letters in stating a rule does not
imply that the formulas designated by those letters must be different. In the &I
rule, for example, @ and Y can stand for any formulas without restriction-even
for the same formula. The following proof of the trivial but valid sequent 'p F
P & P' illustrates this poinr:
1.P A
2. P6cP 1,16.I
86 CuapreR 4
Here we apply the rule of &I (from @ and Y, infer (O & Y)) to a case in which
both @ and Y are'P'. That is, we infer from'P'and'P'again the conclusion'P &
P'. Since we have used 'P' twice, we list line 1 twice in the annotation. Though
odd, this sort of move is quite legitimate, not only for &I but (where applicable)
for other rules as well. Given, for example, that the sun is hot, it validly follows
that the sun is hot and the sun is hot-though we are not likely to have much use
for that conclusion.
The elimination and introduction rules for the biconditional are closely re-
lated to those for conjunction. This is not surprising since 'P * Q' has the same
truth conditions as the conjunction '(P - Q) 6a (Q - P)'. Thus, Iike the conjunc-
don rules, the biconditional rules simply break the complex formula into its con-
ditional components or assemble it from these components:
Biconditional Elimination (*E) From (<D * Y), infer either (<D
- Y) or
(Y * O).
P
↑ ¨
↓ ↓ ′ヽ
A
P
2
A
P
3
1*E
^い︶
Q
4
2r3 -E
kl
P
5
2,4
'We
"eliminate" the biconditional at line 3, obtaining one of its component condi-
donals, 'P * Q'. Next we use modus ponens (-E) at line 4 to obtain'Q', one of
the conjuncts of our desired conclusion. The other coniunct, 'P', was already given
as an assumption. Conjunction introduction enables us to combine these coniuncts
into our conclusion at line 5.
The following proof of '(P Q) * P), P
- (Q-
QFP Q' illustrates the
- *
use of the other biconditional rule, biconditional introduction:
1. (P*Q)*(Q*P) A
2. p-e A
3. Q-P 1,2-E
4. P-Q 2,3*l
'We next consider the disjunction introduction rule (sometimes called the
addition rule):
Disjunction Introduction (vI) From @, infer either (O v Y) or (Y v @).
That is, given any formula, we may infer its disiunction (as either first or second
disjunct) with any other formula. If, for example, my best friend is Jim, then it is
Crassrcar PRoposrrroruaL Locrc: lrurrneruct 87
certainly true that either my best friend is Jim or my best friend is Sally. And it
is obvious from the valuation rule for 'v' that this pattern is valid in general,
for whenever either disjunct of any disjunction is true, the disjunction itself is
also true.
The following proof of '{P-v-Q)--4r3 F R v S' illustrates the use of vI:
1
(PvQ)-P A
P P R R
2
A
3
V Q 2vl
4
1.,3 *E
5
∨S 4vl
To use the conditional assumption'(P v Q)
- R'we must "eliminate" the condi-
tional by -E. But to do this we must first obtain its antecedent, 'P v Q'. Since we
are given'P' as an assumption, we can infer'P v Q'simply by applying vI at line
3. This enables us to derive 'R' at line 4. The conclusion we want to reach, how-
ever, is 'R v S'. But this can be deduced from 'R' by applying vI once again, this
time to line 4.
The disjunction elimination rule, vE, allows us to draw conclusions from
disjunctive premises, provided that we have established certain conditionals:
Disiunction Elimination (vE) From (<D v Y), (@
- @), and (Y
- @),
infer @.
T T
T T
T
F T
F T
F T
F T
F
T T
T
F T
T
T
F
T
F
F
T
T
F F
F F
F F
T T
T
F
1. RvP A
88 CunpreR 4
2.R→ Q A
3.P→ Q A
4.Q l,2,3∨ E
Hcrc Φ is`R',Ψ is`P',and O is`Q'.Noticc that sincc∨ E uscs thrce premises,we
nlust cite three lines to the right when using ito Son■ etiines the sarne line is cited
tWiCC,aS in the pr00f Of`P∨ RP― )Q卜 Q':
1.P∨ P A
2.P→ Q A
3.Q l,2,2∨ E
In thiS pr00f,Φ and Ψ are bOth`P'and O iS`Q'.ThC ttSt prCmiSe iS,Of COurSe,
redllndant,but redundancy docs not affcct validity.
T'he nlost interestipg uses of∨ E are thosc in、 vhich the conditional prenlises
neccssary for proving the concluslon are not given as assumptions but nlustthen■ ―
selvcs be provcd.・ his,ho、 vcvcち rcquires the use of the rule― →I,which is intrO―
I・
・I.his rulc,too,is obviously valid.For by thc valuation rulc for`∼ ',if∼ ∼Φ is truc,
then∼ Φ is falsc and hence Φ is true.TO saゝ fOr exainplc,that l ain nOt nOt tired is
thc sanle thing as to say that l anl tired.I‐ Iere is an exanlple ofthe use of negation
eliminatiOn,in the pr00f Of`P→ ∼∼Q,P卜 Q':
1.P→ ∼∼Q A
2.P A
3.∼ ∼ Q l,2→ E
4. Q 3∼ E
Neither the llegation eliinination rule nor any ofthc othcr rulcs a110w us tO operate
insidc formulas.It is a mistakc,for cxamplc,to do the proof just illustrated this
vvay:.
1.P→ ∼∼Q A
2.P A
3.P― )Q l∼ E(wrong!)
4.Q 2,3-E
Ncgation elimination operates only on doubly ncgatcd formulas.`P― →∼∼Q'is a
collditional,1lot a doubly negated fornlula.NVie nlust use conditional elil■ lillatiOn
tO SCparatC`∼ ∼Q'(WhiCh′ S a dOubly negated fOrmula)■ Omthe COnditiOnal befOre
ncgatiOn cliininatiOn can bc aiplplicd.
It iS nOt really inValid tO elinlinate dOuble negatiOnS inSide fOrnlulaS;it'S iuSt
not a lcgitilnatc usc of our ncgation clilnination rule.ヽ Vie never need to use it this
v′ ay, because our eliinination rules always cnablc us to brcak forinulas down
Cmssrcal PRoposrrrolrar Locrc: llrrrRerucr 89
(where this may validly be done) so that the double negation sooner or later
appears on a line by itself and hence becomes accessible to the negation elimination
'We
rule. could be more liberal, permitting elimination of double negations inside
formulas, but only at the expense of complicating some of our metatheoretic work
later on. Conservatism now will pay off later.
Finally, we should note that there is no one correct way to prove a sequent.
If the sequent is valid, then it will have many different proofs, all of them correcr,
but varying in the kinds of rules used or in their order of application. Often,
however, there is one simplest proof, more obvious than all the rest. In construct-
ing proofs, good logicians strive for simplicity and elegance and thus make their
discipline an art.
Exercise 4.2
Construct proofs for the following sequents:
We have now encountered eight of the ten rules. I saved the remaining two until
last because they make use of a special mechanism: the hypothetical derivation. A
hypothetical derivation is a proof made on the basis of a temporary assumption,
or hypothesis, which we do not assert to be true, but only suppose for the sake of
90 Cunprrn 4
Suppose we stay an extra day at the lake. Then we would get home on
Sunday. But then it would be hard to get ready for school on Monday.
Here the arguer is not asserting that she and her audience will stay an extra day at
the lake, but is only supposing this to see what follows. The conclusion, that it will
be hard to get ready for school on Monday, is likewise not asserted or believed.
The point is simply that this conclusion would be true f the hypothetical supposi-
tion were true.
Her reasoning presupposes t'rvo unstated assumptions, used respectively to
derive the second and third sentences. These are
and
2. If we get home on Sunday, then it will be hard to get ready for school on
Monday.
Using 'S' for 'we stay an extra day at the lake,' 'H' for 'we get home on
Sunday', and 'M' for 'it will be hard to get ready for school on Monday', we may
formalize this reasoning as follows:
1
S*H A
2
H-M A
S H M
3
H (for -I)
4
L,3 -E
5
214*E
(Assumptions 1 and 2 correspond to the implicit statements 1 andZ above. State-
ments 3, 4, and 5 represent the first, second, and third sentences of the stated
argument, respectively. )
I have done something novel beginning with S on line 3, the line that repre-
sents the supposition or hypothesis that we stay an extra day at the lake. Instead
of labeling S as an assumption ('A'), I have marked it with the notation 'H (for
-*I)'. This indicates that 'S' is a hypothesis ('H'), made only for the sake of a
conditional introduction ('--I') argument and not (like 1 and 2) really assumed
and asserted to be true. Moreover, I have drawn a vertical line to the left of 'S'
extending to all subsequent conclusions derived from 'S'. This line specifies that
the reasoning to its right is hypothetical-that statements 3, 4, and 5 are not
genuinely asserted, but only considered for the sake of argument.
This hypothetical reasoning has a purpose. In granting assumptions 1and2,
we see that we can derive 'M' from 'S'; this means the conditional 'S - M' must
be true. This conditional, which symbolizes the English sentence 'if we stay an
extra day at the lake, then it will be hard to get ready for school on Monday', is
both the point of the argument and its implicit final conclusion. But this condi-
Crasstcal PRoposrrroruar Locrc: lNrrRrncr 91
tional is not deduced directly from our assumptions, nor from any of the state-
ments listed in the argument, either singly or in combination. Rather, we know
that'S -- M' is true because (given our assumptions) we showed in the hypotheti-
cal reasoning (or hypothetical derivation) carried out in lines 3-5 that 'M' follows
logically from 'S'. It is this reasoning, not any single statement or set of statements,
that shows 'S - M' is true. To indicate this, and to draw the argument's final
conclusion, we add a new line to the previous reasoning, as follows:
1.S― →H A
2
H→ M A
3
S H(fOr→ I)
4
H 1,3-→ E
5
M 2,4-→ E
6
S→ M 3-5-→ I
The annotation of line 6 indicates that we have drawn the conclusion 'S * M'
from the hypothetical derivation displayed on lines 3-5. The rule used is the rule
of conditional introduction (-11, commonly known as conditional proof. It may
be stated as follows:
A
Q H(fOr→ I)
P&Q 1,2&I
・
Q
→ (P&Q)
4
2-3-→ I
92 Cnaprrn 4
The sequent's conclusion, 'Q * (P & Q)', is a conditional, so after listing the
assumption 'P' as usual, we hypothesize the antecedent 'Q' of this conditional at
line2. A single step of &I at line 3 enables us to derive its consequent,'P & Q',
thus completing the hypothetical derivation. We then get the desired conclusion
by applying conditional introduction to the hypothetical derivation atline 4.
Conditional introduction is used in proving biconditional conclusions as well
as conditional conclusions. But in proving biconditionals we often need to employ
it twice in order to prove each of the two conditionals that comprise the bicondi-
tional before we assemble these components into the biconditional conclusion.
The following proof of the sequent 'P 6c Q F P * Q' illustrates this technique:
1. P&Q A
2. 1 P H(for-I)
3.r a l&E
4. P*Q 2-3-l
5. 1 a H(for-I)
6. I P l&E
7.Q-P 5-6-I
8. P*Q 4,7*l
Here the conclusion we wish to obtain is 'P * Q'. The rule for proving bicondi-
tional conclusions is *I, but to use *I to get 'P * Q' we must first obtain its
"component" conditionals, 'P
- Q' and 'Q -* P'. We do this in lines 2-4 and 5-
7, respectively, by first hypothesizing each conditional's antecedent, next hypo-
thetically deriving its consequent (which in each case involves a simple step of &E
from our assumption), and finally applying *I to the resulting hypothetical deri-
vation (at lines 4 and 7, respectively). Having obtained the two component con-
ditionals, we complete the proof with a step of -I at line 8.
A step or two of conditional introduction is often used to provide the condi-
tionals needed for drawing conclusions from a disjunctive premise by vE. This
proof of 'P v P l- P' provides an example that is both elegant and instructive:
1. PvP A
2. I P H(for-I)
3. P-P 2-2-I
4. P 1,3,3 vE
Our assumption is the disjunctive premise'P v P'. The standard rule for drawing
conclusions from disjunctive premises is vE: From (O v Y), (@- @), and (Y- O),
infer O. If we take @, Y, and @ all to be 'P', this becomes: From 'P v P', 'P
- P',
and 'P * P', infer 'P'. Thus we see that if we can prove 'P * P', we can use it twice
with our assumption 'P v P' to deduce the desired conclusion 'P'. But how do we
prove 'P * P'? That's where --I comes in. rWe hypothesize this conditional's ante-
cedent at line 2 and aim to derive its consequent. The hypothetical derivation is
the simplest possible, for its hypothesis and conclusion are the very same statement
'P'. There is no need to apply any rules. In hypothesizing'P',we have already in
effect concluded 'P'; the hypothetical derivation ends as soon as it begins atline2.
'We
then use *I to derive'P
- P' at line 3 and vE to obtain 'P' at line 4.
Crassrcnr PRopostttoruaL Loctc: lNrrRrrcr 93
Let's consider one more example of the use of -I in preparation for a step of
vE. In this case the sequent to be proved is 'P v Q, R F (P & R) v (Q & R)':
1.PV Q A
2.R , A
2
υ 。
1
1
D■ H(fOr→ I)
4.I P8こ R 2,3 8cI
5。 │ (P&R)χ (Q&R) 4 vI
6.P→ ((P&R)∨ (Q&R)) 3-5-→ I
7。 I Q H(fOr→ I)
8。 I Q&R 2,7 8cI
9。 │ (う &R)∨ (Q&R) 8∨ I
lo.Q― ((P&R)∨ (Qに R))' 7-9-→ I
11。 (P&R)∨ (Q&R) l,6,10∨ E
To use the disjunctive premise 'P ., Q'to obtain the conclusion'(P & R) v (Q &
R)' by vE, we need two conditional premises: 'P -* ((P & R) v (Q & R))' and
'Q -. ((P & R) v (Q & R))'. These are conditionals, so we use -*I to prove each,
thl first in lines 3-6, the second in lines 7-10. Once the two conditionals have
been established, a single step of vE at line 11 completes the proof.
In proving conditionals whose antecedents contain further conditionals, we
sometimes need to make two or more hypothetical suppositions in succession. For
,P (Q -+ R) -+ (Q * (P 6a R))" we hypothesize the conclusion's
example, to prove l-
'Q * R' and then aim to deduce its consequent'Q * (P & R)'. But this
"rt...d..rt
consequent is itself a conditional so that we must introduce a second hypothesis,
the second conditional's antecedent, 'Q'. This enables us to deduce 'Q * (P & R)'
by -.I. And since this is proved under the initial hypothesis '(Q -- R)', a final step
oi *I yields the conclusion '(Q R) (Q -* (P & R))'. Here is the proof in full:
- -
1. P A
2. Q→ R H(br→ I)
3. Q H(fOr→ I)
4. R 2,3-→ E
5。 P&R l,4&I
6. Q→ (P&R) 3-5-→ I
7. (Q-*R)-(Q*(P&R)) 2-6-I
Notice that though the antecedent of '(Q - R) - (Q * (P & R))' is also a
conditional, 'Q .* R', we do not attempt to prove this conditional by hypothesiz-
irg 'Q' and deriving 'R'. The antecedent of a conditional conclusion, no matter
how complex, typically figures in a proof as a single hypothesis (line 2 in the proof
above) and is not itself proved.
Finalln after a hypothetical deriuation ends, all the formulas contained
*off limits"
within it ar:e for the rest of the proof. They may not be used or cited
later, because they were never genuinely asserted, but only hypothetically enter-
tained. The following attempted proof of the invalid sequent 'R Q * -P F P & -P'
illustrates how violations of this restriction breed trouble. (If you don't see that
this sequent is invalid, check it with a truth table.)
94 CHaprrn 4
1. P A
2.Q→ ∼P A
3.I Q H(br→ I)
4. │ ∼P 2,3-→ E
5.Q→ ∼P 3-4-→ I
6。 P8こ ∼P l,48こ I(WrOng!)
All rules are used correctly through step 5, though steps 3-5 are redundant, since
all they do is prove 'Q -. -P', which was already given as an assumption at line 2.
Step 6, however, is mistaken, since it uses the formula '-p', which appears in the
hypothetical derivation at line 4, after that hypothetical derivation has ended. '-p',
however, was never proved; it was merely derived from the supposition of 'e'. It
cannot be cited after the hypothetical derivation based on 'Q' ends at step 4.
violation of this restriction may result in "proofs" of invalid sequents, as it does
here. These, of course, are not really proofs, since in a proof the rules must be
applied correctly.
However, any nonhypothetical assumption or nonhypothetical conclusion
and any bypotbesis or conclusion within a bypothetical deriuation tbat has not yet
ended may be used to draw further conclusions. So, for example, in the proof of
'P l- (Q -* R) * (Q * (P & R))', which was given just before the preceding
example, it is permissible to use the hypothesis 'Q
- R' (line 2) at line 4 of thi
hypothetical derivation that begins with 'Q' (line 3), because the hypothetical
derivation beginning with'Q -t R' has not yet ended.
A proof is not complete until all hypothetical derivations have ended. If we
were to leave a hypothetical derivation incomplere, then its hypothesis would be
an additional assumption in the reasoning; but, being marked with an 'H'instead
of an '/t', it might not be recognized as such.
To summarize:
-I is the rule most often used for proving conditional conclu-
sions. To prove a conditional conclusion @ * Y, hypothesize its antecedent @ and
reason hypothetically to irs consequent Y. Then, citing this entire hypothetical
derivation, deduce (D .* Y by *1. The conclusion O -. Y does not belong to the
hypothetical derivation, so the vertical line that began with @ does not continue
to (D - Y, but ends with Y.
It is perhaps not so obvious as with the nonhypothetical rules that is
valid. To recognize its validitS we must keep in mind that the hypothetical -Ideri-
vation from @ to Y must itself have been constructed using valid rules. This means
that if a valuation makes true both the proof's assumptions and <D, as well as any
other hypotheses whose derivations had not ended when (D was supposed, then it
also makes Y true. That is, there is no valuation that makes these assumptions and
hypotheses true and also makes o true but Y untrue. In other words, ihere is no
valuation that makes these assumptions and hypotheses true and (D .- y untrue.l
But this means that the inference from these assumptions or hypotheses to (D * y
is valid. Hence the rule
-I, which allows us to conclude <D * y from these as-
1 This reasoning appeals implicitly to the valuation rule for the conditional.
Crassrcnr PRopostttonnr Loctc: lNre RrNcr 95
sunlptiOns and hypothescs,is itself valid;it never icads fronl truc prenlises to an
untrue cOnclusion.
N▼ ic next consider thc rulc for proving ncgativc propOSitions:negation intro―
duction,∼ I,。 ften knOwn as indirect proof or“ θご%ε″ο αJ αbSZ名グ%″t(reduCtiOn tO
absurdity).Ncgatioll introduction is the rulc for proving negatcd conclusions.'T10
prOvc∼ Φ,hypothesize Φ and validly derive fronl Φ an``absurdity"― that is,a
conclusion known to be false.Since the derivation is valid,if Φ and any additional
assunlptions or hypothescs used in the dcrivation、 verc truc,the dcrived conclusion
would have to be true as well.Therefore,since the derivcd conclusiOn is false,
cither Φ or sorne other assunlption Or hypothesis use(l to derive it inust bc falSC.
SO,if these other assunlptions or llypotheses arc truc,it nlust be Φ that is falsc.
Hencc∼ Φ f01lo、 vs fronl these other assunlptions or llypothcses.
But how can we forinally cnsurc that thc conclusion wc derivc from Φ is
falscP One way is to requirc that the conclusion be inconsistcnt.Inconsistencies of
the fOrin`Φ 8こ ∼Φ',fOr exarnple,11l the bill.Actually,any inconsistcncy、 vould do,
but sO as l10t tO ullduly conlplicate our rule,、ve sllall requirc that the conclusion
Of the hypothetical dcrivation always havc this onc fornl.This restriction,as we
sllall sce in Section 5.10,does not iprcvent us fronl proving any Valid sequent.
T'hcreforc we will statc the negation introduction fulc as follows:
A
2
A
3
P. H(fOr∼ I)
4
Q 1,3-→ E
5
Q&∼ Q 2,48こ I
一
6
3-5∼ I
Having listed the assumptions on lines 1. and2,we note that the desired conclusion
is a negation,
,-P,. To prove this conclusion by -I, then, we hypothesize 'P' at line
3-noi, as before, for .+I, but rather for -I-and try to derive an "absurdity."
This is accomplished at line 5, where it is established that, given the assumptions
'P - Q' and '-Q', 'P' leads to absurdity. Therefore, given these assumptions, 'P'
must be false, which is what we conclude at line 6 by asserting '-P'.
Formal indirect proofs are, of course, not merely formal. They may be used
to represent specific natural language arguments. So, for example, if we let 'P'
,trrrd for ,A pirson is defined by her genome' and 'Q' for 'Identical twins are the
same personl the reaso.ring represented by this proof is as follows. It is assumed
at lint 1 that if a person is defined by her genome, then identical twins are the same
person and at line 2 that identical twins are not the same person. The argument
to show that a person is not defined by her genome (line 6). To prove this, we
"i-,
suppose for the sake of argument at line 3 that a person ls defined by her genome.
W. do not, of course, really assert this; we suppose it only to reduce it to absurdity
95 CHaprrn 4
and so prove its negation. Together with assumption 1, this supposition leads at
line 4 to the conclusion that identical twins are the same person. And this conclu-
sion, together with assumption2, yields the absurd conclusion that identical twins
both are and are not the same person. Having shown, given assumptions 1- and2,
that the supposition rhar a person is defined by her genome leads to absurditr we
conclude on the strength of these assumptions alone that a person is not defined
by her genome. This final conclusion is recorded on line 5.
The following proof of the sequent '-(P v Q) F -p' provides another example
of the application of -I. Recall that '-(P v Q)' means "neither p nor e.,,
1. -(Pv Q) A
2. 1 P H(for-I)
3. I P'Q 2vr
4. I (P., Q) & -(P v Q) 1,3 &I
5. -P 2_4 -r
'sfith
respect to our statement of the negation introduction rule, @ here is ,p' and
Y is 'P v Q'. once again the conclusion to be proved is '-p'. So, after listing the
assumption, we hypothesize 'P' and aim for some contradiction. The trick is to see
that we can obtain 'P r Q', which contradicts our assumption, by applying vI to
'P'. The contradiction (absurdity) is reached at line 4 by &I. 'p' having led to an
absurdity, we deduce '-P' at line 5.
Negation introduction may also be used, in combination with negation elim-
ination, to prove unnegated conclusions. To prove an unnegated conclusion o, we
may hypothesize -o, derive an absurdity, and apply -L But since -I adds a nega-
tion sign to the hypothesis that is reduced to absurdity, it enables us to conclude
only --@, not the desired conclusion @. However, from --@ we can deduce @ by
negation elimination and so complete the proof. The following proof of '-(p & -e),
P F Q' uses this straregy.2 In this case o is 'Q'; with respect to the formal statemenr
of the -I rule, (D it '-Q' and Y is 'P & -Q':
1・ ∼(P&∼ Q) A
2
P A
3
-a H (for -I)
4
P&-Q 2,3kr
5
一
”
3-5 -I
´ヽ
ヽこ
7
6-E
Negation introduction is often combined with conditional introduction, as
in this proof of the sequent, 'P
inference called contraposition:
- Q F -Q * -P', which expresses the pattern of
1. p--e A
2. I -a H(for-I)
2 To see why this form ought to be valid, recall
that '-(P & -e),is equivalenr to
.p* e,.
Crassrcar PRopostrtoruaL Loctc: lrurearrucr 97
3. P H(fOr∼ I)
4. Q 1,3-→ E
5。 Q&∼ Q 2,4 8ζ I
6。 ∼P 3-5∼ I
7.∼ Q→ ∼P 2-6-→ I
Having written our assumption, we note that the conclusion for which we are
aiming, '-Q r -P', is a conditional. So we hypothesize its antecedent at line 2 for
'-P'. But '-P' is a negation, and -I is the rule
-I, aiming to derive its consequent,
for proving negations. So, to set up a derivation of '-P', we hypothesize 'P' at line
3 for -I and try to deduce a contradiction. The contradiction is obtained at line 5,
which enables us to use now derived '-P' from'-Q',
-I at line 6 to get '-P'. Having
the proof.
we can deduce '-P - -Q' by -Lt line 7 to complete
Negation introduction is used in a peculiar way in the proof of the principle
ex falso quodlibet, the principle expressed by the sequent 'P, -P F Q'. (We demon-
strated the validity of this sequent using a truth table in Section 3.2.)
P ”
1
A
2
A
3
∼Q H(br∼ I)
4
P8こ ∼P 1,2&I
Q
5
3-4∼ I
一
´ヽ
一
6
5∼ E
ヽ こ
Q Q
■
P
ヽ
A
′ ↓ 一
2
P
¨ ︱
A
■
P H(fOr→ I)
4
∼Q H(fOr∼ I)
,
P8こ ∼P 2,38こ I
Q
‘
一
4-5∼ I
Q
Z
一
,
6∼ E
&
P ︲
3-7-→ I
Q
,
H(br→ I)
Q
C
,
と
9,9-→ I
t
C
を
l,8,10∨ E
our first assumption is a disjunction; to use it we need vE. But to use vE with
'Pt Q' to obtain the conclusion 'Q', we need these two conditionals: .p * e, and
'Q - Q'. These we obtain by
To prove * -I, the first in lines 3-8, the second in lines 9-10.
'P Q', we hypothesize its antecedent 'p' at line 3. We now have
hypothesized 'P' and assumed '-P' so that we can obtain any conclusion we please.
'we want 'Q', the consequent
of 'P - Q', in order to complete our conditional
proof. To get it, we hypothesize '-Q' for reduction to absurdity. As in the previous
example, however, we derive the absurdity (at line 5), not from this hypothesis but
from previous (and irrelevant) assumptions. Nevertheless, this allows us to con-
clude'--Q'at line 5by -1, from which we obrain'Q'at line 7.The hypothetical
derivation at lines 3-7 has thus established 'P * e', d fact we record at line g. The
proof of 'Q * Q' at lines 9-10 is trivial. Having obtained the necessary premises
at lines 1, 8, and 10, we finish with a step of vE.
Although there are many (indeed, infinitely many!) different proofs for each
valid sequent, there is often one way that is the simplest and most direct. Finding
that way is a matter of strategy. often the best strategy for a proof can be "readi
directly from the form of the conclusion-that is, from the identity of its main
operator, as Table 4.1 indicates.
It is common, as we have seen in some of the examples worked earlier,
for different strategies to be used successively in different stages of a proof. To
illustrate how Table 4.1 provides guidance in doing this, let's prove the sequenr
'P r Q F Q v P'. we begin by noting that the conclusion of this sequent is of the
form <D v Y. The first suggestion in the table for conclusions of this form is to use
vI if either (D or Y (i.e., in this instance 'P' or 'Q') is present as a premise. But we
have neither premise, so this suggestion is inapplicable. 'we then try the second
suggestion, which is applicable if there is a premise of the form @ v A. ,p v e'
is such a premise. The table then recommends proving as subconclusions the
conditionals @* (@ v Y) and A- (O v Y) (i.e., in this case,p* (e r p)' and
'Q * (Q v P)'). A subconclusion is simply a conclusion useful for obtaining the
main conclusion. It may be, but is not always, the conclusion of a hypothetical
derivation.
Now the task is to prove the two subconclusions 'P * (Q v P), and .q *
(Q P)'. These are both of the form o -* Y. So we consult Table 4.1 regarding
'
strategies for proving conclusions of this form. The table recommends in each casi
Cnssrcar Pnopostttoruau Loctc: lNrrnrlce 99
TABLE 4.1
Proof Strategies
Φ〔
蛯Ψ Prove the subconclusions o and Y separately and
then join them by &I.
Φ ∨Ψ If either (D or Y is a premise, simply apply vI to
obtain O v Y. Otherwise, if there is a disiunctive
premise @ v A, try proving the two conditionals
O * v Y) and A - ((D v Y) as subconclusions and
(@
then using vE to obtain @ v Y. If neither of these
strategies works, then hypothesize -(@ v Y) and
work toward a subconclusion of the form @ & -@ in
order to obtain @ v Y by -I and -E.
'We
to hypothesize the antecedent, derive the consequent, and then use
-I. are
now working on two levels. Though our ultimate strategy is to use vE, our imme-
diate strategy is to use *I to obtain both 'P * (Q v P)' and 'Q * (Q v P)'. We
begin the proof of the first of these conditionals by hypothesizing 'P' (line 2 in the
proof below).
Now our goal is to prove yet another subconclusion: 'Q v P'. This, again, is
of the form @ v Y. So once again we consult Table 4.1 for strategies for proving
conclusions of this form. But now, since we have hypothesized 'P" we can follow
the first suggestion for disjunctive conclusions, obtaining 'Q t P' directly by vI.
This enables us to prove 'P - (Q v P)' by -I, completing our -I strategY.
The second conditional, 'Q .* (Q v P)', can now be proved in the same way (lines
5-7 below). Now with both conditionals available, we simply apply vE (at line 8
below), completing both our initial strategy and the proof. Here is the result:
1
P∨ Q A
P Q ∨Q Q
2
H(fOr→ I)
Vい
P P
3
2∨ I
Q
4
刷.
pニ
2-3-→ I
V
5
H(fOr→ I)
6
5∨ I
100 Craprrn 4
Though the form of the conclusion usually determines the best overall strat-
egy for the proof, the details depend on rhe forms of the premises. It is almost
always necessary to break the premises into their components using the appropri-
ate elimination rules. Sometimes when no promising strategy is apparent, breaking
down the premises in this way makes the path to the conclusion clear. As the
preceding example illustrates, disjunctive premises typically demand a disjunction
elimination strategy. where (D v Y is the disjunctive premise and @ is the desired
conclusion, this strategy requires the conditionals (D * @ and Y * @, which must
usually be proved by *I before vE can be applied.
Sometimes none of the strategies listed in Table 4.1 seems to work. In that
case, consider what additional premises would enable you to prove the conclusion,
and see if these can be proved in some way. In many cases, such premises can be
proved by -I.
The following proof of '-P Q F P v Q' illustrates this point:
-
1
∼P→ Q A
2
∼(PVQ) H(br∼ I)
3
P H(fOr∼ I)
4
P∨ Q 2∨ I
5
∼P 3-5∼ I
7
Q l,6-→ E
8
P∨ Q 7∨ I
9
∼∼(P∨ Q) 2-9∼ I
11・ P∨ Q 10∼ E
Having written the assumption at line 1., we note that the conclusion is of the form
(D v Y and consult Table 4.1 regarding conclusions of this form. But we are not
given either disjunct as a premise, nor are we given a disjunctive premise, so the
first two strategies for disiunctive conclusions are inapplicable. We therefore fol-
low the third, hypothesizing the negation of this conclusion at line 2 and hoping
to derive an absurdity. But which absurdity? Thinking ahead, we see that if we
had the premise '-P', then by
-E with 1 we could get'Q', from which we could
obtain 'P v Q' by vI. This would contradicr our hypothesis, providing the desired
absurdity. Hence we can solve the problem if we can prove '-P'. Since '-P' is of the
form -(D, Table 4.1 recommends using -I as the strategy for proving it. So we
hypothesize 'P' at line 3, attempting to derive an absurdity. This is achieved at line
5, which enables us to conclude '-P' at line 6. The rest of the proof then proceeds
according to the strategy just outlined. Solving difficult problems in propositional
logic often requires just this sort of thinking ahead.
This proof of 'P * -Q F -(P * Q)' uses a similar double -I strategy, which
requires the same sort of advance planning:
Cussrcar PnopostttoruaL Loctc: lurrRrruce 101
1.P⇔ ∼Q A
2. P→ Q H(fOr∼ I)
3. P→ ∼Q 1+→ E
4. ∼Q→ P l→ E
5。 P→ Q 2-E
6. Q→ P 2-E
7. Q H(fOr∼ I)
8. P 6,7-→ E
9. ∼Q 3,8-→ E
10. Q&∼ Q 7,9 8cI
11. ∼Q 7-10∼ I
12. P 4,11-→ E
13. Q 5,12-→ E
14. Q&∼ Q ll,13&I
15。 ∼(P→ Q) 2-14∼ I
Since the sequent's conclusion is '-(P * Q)', which has the form -@, we hypothe-
size 'P * Q' at line 2 in order to reduce it to absurdity. No strategy is apparent at
this point, so we use *E to break the two biconditionals into their components at
Iines3-5, hoping that this will help us see how the absurdity might be derived.
And does, for we now note that if we could prove '-Q', then using line 4 we
it
could get'P', and with 'P'together with line 5 we could get 'Q'. Then we would
have both 'Q' and '-Q'-which would provide the absurdity that we need to
derive from the hypothesis 'P * Q'. But before we can do any of this, we must
prove Since this is a negated formula, the appropriate strategy is -I. So we
'-Q'.
hypothesize 'Q' at line 7 in order to reduce it to absurdity. We obtain the absurdity
10, and that yields'-Q'at line 11. We then proceed to the final absurdity
"iiirr.
as planned.
The ten rules of inference are summarizedinTable 4.2.
Exercise 4.3
Construct a proof for each of the following sequents:
1.P→ Q,P卜 Q∨ R
2.(PvQ)→ R卜 P→ R
3.P→ Q,P→ R卜 P→ (Q&R)
4.P→ Q卜 (P&R)→ Q
5。 P⇔ Q,Q→ R卜 P→ R
6.P→ (Q→R)卜 (P&Q)→ R
7.(P&Q)→ R卜 P→ (Q→ R)
8。 P→ Q卜 (Q→ R)→ (P→ R)
9。 P卜 Q→ (R→ (S→ ((P&Q)&(R&S))))
10。 ∼∼P∨ ∼∼Q卜 P∨ Q
ll.(P&Q)∨ (P&R)卜 P
12。 (P&Q)∨ (P&R)卜 Q∨ R
13.P→ Q,P→ ∼Q卜 ∼P
102 Cuapren 4
TABLE 4.2
Summary:The Tlen Ru:es ofinference
Coniunction Introduction (Confunction) (&I) From @ and Y, infer (<D & Y).
Disjunction Elimination (Constructive Dilemma) (vE) From (tD v Y), (<D- @), and
(Y O), infer @.
-
Disiunction Introduction (Addition) (vI) From @, infer either (@ v Y) or (Y v @).
14. P― →∼P卜 ∼P
15。 P,∼ Q卜∼(P→ Q)
16.∼ (P&Q)卜 P→ ∼Q
17。 P→ Q卜 ∼P― ∼Q
18。 (P&∼ Q)∨ (∼ P&Q)卜 ∼(P→ Q)
19・ P∨ Q卜 ∼(∼ P∨
∼Q)
20。 ∼P卜 P→ Q
A ren■ arkable feature Of thc hypOthetical rulcs is that thcy cnable us to prove
conclusiOns withOut assunlptiOns.(〕 onsidcち for exanlplc,this sinlplc pr00f Of thc
assunlptiOnless sequent`卜 P… →P':
1。 I P H(fOr‐ →I)
2. P― →P l-1-→ I
But this is reasOnable,fOr`P― →P'is a taut010gy一 that is,a valid fornlula.It is truc
on all valuatiOns― that is,nO nlatter what thc facts arc一 and hence is truc rcgard‐
lCSS Of What we assume,or whether wc assume anything at all.Each of the for―
Cmssrcal PRopostttoruaL Loctc: lNrrnrnce 103
mula's instances-the statement 'lf the pie is done, then the pie is done', for
example-is true come what may. (Even in a world where there are no pies, it is
true that if the pie is done, then the pie is done!) Conclusions provable without
assumptions are called theorems. And, as we shall see in Section 5.10, all the
theorems of propositional logic are tautologies, and vice versa.3
.-(P &
The foimula -P)', which expresses the principle of noncontradiction,
is also a theorem, which is equally easy to prove:
1. I P&-P H(for-I)
2. -(P E -P) 1-1 -I
Here we needn't draw any conclusions in order to reduce the hypothesis to absurd-
ity; the hypothesis ls an absurdity!
Sfighlly more difficult to prove than these theorems is the law of excluded
middle as expressed by the sequent'F P v -P':
∼(PV∼ P) H(fOr∼ I)
3
P∨ ∼P
4
妥
6
P∨ ∼P 5 vIプ
7
P∨ ∼P 8∼ E
Here we prove 'P v -P' by reducing its negation, '-(P v -P)', to absurdity. The
trick is to see, after hypothesizing '-(P v -P)', that we could contradict this hy-
pothesis by a simple step of vl (line 6) if we could prove '-P'. But '-P' is a negated
iormula, so we hypothesize 'P' at line 3 in an effort to reduce it to absurdity. This
absurdity emerges at line 4, enabling us to deduce '-P' at line 5 and so complete
the strategy as planned.
As a final example, we shall prove the theorem 'F (P - Q) - (-Q * -P)"
The strategy here, an indirect proof nested within two conditional proofs, calls for
three hypotheses:
1
P→ Q H(fOr→ I)
2
∼Q H(fOr→ I)
3
P H(fOr∼ I)
4
Q 1,3-→ E
5
∼P 3-5∼ I
7
∼Q→ ∼P 2-6-→ I
8
'When
we prove a sequent, whether or not it is a theorem, we show it to be a
reliably valid inference pattern-a fact which can be used to shorten proofs. Con-
sider, for example, the sequent 'P l- --P', which expresses an inference pattern that
like -E is often called double negation. Its proof is as follows:
1 ェ ●る ●
A
I
-P H (for-I)
j ´
︲ ”
P&-P 1.,2 Ur
p■
4
1-2 -l
Having proved this sequent, we have in effect proved the validity of all sequents
of the form (D F --@, where @ is any formula whatsoever-sequents rr.h
"t
'P r Q l- --(P v Q)', 'P .* -R l- --(P * -R)', '-Q F ---e,, and so on. W. shall call
these uarian s of the original sequent. More precisely, a variant of a sequent is a
sequent formed by replacing one or more of its sentence letters by formulas, each
occurrence of the same senrence letter being replaced by the same formul a. (Note,
ltoweuer, that only sentence letters may be replaced, not larger parts of the for-
mula.)Take, for example, 'P r Q f- --(P v Q)?. This is a variant of the sequent ,p F
--P' because it is the result of replacing each occurrence of the sentence letter'P'
in 'P F --P' by 'P r Q'. To prove 'P l- --P' is in effect ro prove ,P r e l- --(p v e)'
as well, since the proof of the latter mimics the proof of the former precisely; we
simply replace each occurrence of 'P' in the proof of 'P l- --p' with ,p v e,, leaving
everything else the same:
1. PvQ A
2. | -(P,, Q) H (for -I)
3. I (PvQ)&-(pve) 1,2 Ul
a. --(P v Q) 1-2 -l
Any other variant of 'P F --P' can be proved by the same sort of replacement.
Thus, having proved'P F --P', we have in effect shown how to p.ou.."-h instance
of the general rule of double negation:
1'--Q*R A
2.Q A
3. --Q 2DN
4.R 1,3 *E
Any previously proved sequent may be used as a rule of inference in this way. Such
rules are called derived rules because their justification is derived from the ten rules
we have taken as basic. Derived rules do not enable us to prove any sequent that
cannot akeady be proved by the original ten. They merely shorten proofs and save
work.
Though any previously proved sequent can be used as a derived rule, we
shall be selective in the ones we name and abbreviate for use. Too many rules can
be cumbersome. The derived rules listed in Table 4.3 suffice for an elegant proof
for just about any sequent we might encounter. We have not actually proved them
all-yet. Those that remain unproved are left as exercises.
From now on we may freely use the derived rules listed in Table 4.3. The
following proof of '-(-P v Q) l- -(P Q)' provides a further illustration of the use
-
of derived rules:
1.∼ (∼ P∨ Q) A
) P*Q H (for-I)
3. --PE -Q 1DM
4. -a 3 6cE
5. -P 2,4 MT
6. --P 3 6cE
7. -P 6c --P 5,5 6.I
8。 ∼(P→ Q) 2-7 -r
The overall strategy is reductio (indirect proof). Notice that there is no need to use
-E on line 6 before proceeding to the contradiction. '-P & --P' is iust as good a
contradiction as 'P & -P'.
TABLE 4.3
Some Important Derived Rules
(continued)
106 CHnprrn 4
TABLE 4.3
Some lmportant Derived Rules (contrnued)
Exercise 4.4.1
Sequents representing some of the derived rules discussed in this section have been
proved in the text (Sections 4.2-4.4), but a number of the derived rules have not
been so verified. complete the verification of these derived rules by proving the
following sequents, using only the ten basic rules:
Crassrcar PRoposrrroNar Locrc: lrureRrNce 107
Exercise 4.4.2
Prove the following theorems using either basic or derived rules:
1.卜 P→ (P∨ Q)
2.卜 (P&Q)→ P
3. 卜P― →∼∼P
4.卜 P→ (Q→ (P&Q))
5.卜 P― ((P→ Q)→ Q)
6。 卜∼(P∨ Q)→ ∼P
7. 卜∼(P← →∼P)
8。「 P→ (Q∨ ∼Q)
9。 卜(P→ Q)V(P→ ∼Q)
10。 卜(P→ Q)― ∼(P&∼ Q)
Exercise 4.4.3
Reprove the sequents of Exercise 4.3, using derived rules to shorten the proofs
wherever possible.
108 Crnprun 4
Our system of ten inference rules is only one of many proof systems for classical
predicate logic. All do the same work-that is, prove the same sequents-so
choice among them is largely a matter of taste and style. For the austere, there is
the system which uses only two rules-modus ponens (our *E) and tautology
introduction:
Tautology Introduction (TI) Any tautology may be asserted at any line of
a proof.
limited number of recognizable forms, we can still create a system that will prove
all valid sequents. Such, for example, is the system that has as its inference rules
only *g and the introduction of tautologies of one of the following three forms:
(-P
- -Q)- ((-P
- Q) - P)
by replacing '-P *Q' with 'P v Q'. Definitions such as these are called contextual
definitions or definitions in use, because they give directions for replacing not
merely the defined term, but an entire formula or subformula containing that term,
by an abbreviating formula, and vice versa. The operators'-'and '''
are regarded
in this axiom system as primitive operators; only they are mentioned in the for-
mation rules. Those introduced by definition are known as defined operators.
Formulas containing defined operators are regarded as mere abbreviations for
formulas containing only primitive operators.
The full set of definitions for the defined operators is as follows:
イ
〓
@vY -O*Y
イ イ
o&Y
〓 〓
-((D - -Y)
<D*Y (o*Y)&(Y-.tD)
The symbol '- tf is a metalinguistic abbreviation meaning "is by definition."
Notice that the definition of '*' mentions the operator'&'. It must therefore
be supplemented by the definition of '&' if expressions containing '*' are to be
reduced to primitive terms. Formulas of the form O * Y abbreviate formulas of
the form ((D - Y) & (Y - @), which in turn (by the definition of '&') abbreviate
formulas of the form -((O * Y) - -(Y t O)).
1.Q A
2.Q→ P→ Q) (∼ Al
3.∼ P→ Q l,2→ E
4.P∨ Q 3Df∨
In step 2 we intrOduce an instance of axlom schema l in which Φ is`Q'and Ψis
`∼ P'.1・ he COnCluSiOn fo1lows directly by lnodus ponens(― →E)and thc deinitiOn
of`∨ '.
HCre iS a pr00f Of hypothetical syHogism,`P→ Q,Q→ R卜 P→ R':
1.P→ Q A
2.Q→ R A
3.(P→ (Q→ R))→ ((P→ Q)― (P→ R))Ax2
4。 (Q→ R)→ (P→ (Q→ R)) Axl
5。 P→ (Q→ R) 2,4→ E
6。 (P→ Q)→ (P― R) 3,5→ E
7.P→ R l,6-→ E
In stcp 3 We uSe a simple instancc of axlom schema 2 in which Φ is`P',Ψ is`Q',
and O is`R'.In step 4 wc intrOducc an instancc of axiom scheina l in which Φis
`Q― →R'and Ψ is`P'.The cOnclusion thcn foHows by three stcps ofmodus poncns.
In virtue of this pr00f,hylpOthctical sy1logislll nlay hcncefOrth bc used as a dcrived
rule.
As a inal illustratiOn,wc shall prOvc`∼ ∼P卜 P',which cstablishes∼ E as a
dcrived rule in this ax1011l systcnl:
1. ∼∼P A
2.∼ ∼P‐ )(∼ P一 →∼∼P) Axl
3. ∼P―→∼∼P l,2-→ E
4。 (∼ P―
→∼∼P)―→((∼ P―→∼P)― →P) Ax3
5。 (∼ P―
→∼P)― →P 3,4-→ E
6。 (∼ P―
→((∼ P→ ∼P)― →∼P)一→((∼ P― →(∼ P→ ∼P))―→
(∼ P―
→∼ P)) Ax2
7.∼ P→ ((∼ P→ ∼P)→ ∼P) Axl
8。 (∼ P―〉(∼ P→ ∼P))― → (∼ P→ ― P) 6,7→ E
9. ∼P―→ (∼ P― →∼P) Axl
10. ∼P→ ∼P 8,9-→ E
ll. P 5,10→ E
Classrcar PRopostttorunr Loctc: lxrenrrucr 111
Because this axiom system has just three axiom schemas and only one inference
rule, there is little to work with. Proofs tend to be longer and more complicated
than those in natural deduction systems. This can be alleviated, however, by the
introduction of derived rules. Axiom systems, though perhaps cumbersome to
reason with, are generally easy to reason about. Thus axiom systems are often
preferred for metatheoretic work. But that is the subject of the next chapter.
It can be shown (though we shall not show it) that each of the ten basic rules
of our natural deduction system is a derived rule of this axiom system. Thus any
sequent provable in the former is provable in the latter. Likewise, it is easy to show
that each instance of each axiom of this system is a theorem of our natural deduc-
tion system. Since the natural deductiqn system also contains .*E, this means that
any sequent provable in the axiom system is also provable in the natural deduction
system-provided there is always some way within the natural deduction system
to mimic definitional inferences. This, too, can be shown. Therefore the two sys-
tems prove exactly the same sequents.
Deduction can also be pursued in an algebraic fashion. The idea here is to
treat logical equivalences as identities. Thus, for example, the commutation law,
(O 6( Y)* (Y E O)
may prove a sequent to be valid, but proofs are not designed to reveal invalidity.
If we try to prove a sequent and fail, that does not show the sequent is invalid.
Maybe we did not try hard enough. Thus, given a sequent whose validity is in
question, trying to prove it settles the question only if the sequent is valid and we
do in fact find a proof.
Moreover, because proof rules can be applied or axioms introduced in any
order, they may be applied repeatedly without ever reaching the desired conclu-
sion, even if that conclusion validly follows from the premises. Proof, in other
words, may elude us, even though proofs exist.
Because of these disadvantages, this book does not emphasize proofs, though
they are the most familiar and widely recpgnized way of doing logic. We shall
concentrate instead on the more powerful semantic techniques-valuation rules,
truth tables, and trees-and their generalizations for more advanced logical
systems.
Exercise 4.5.1
Prove instances of Ax1, Ax2, and Ax3 as theorems within our natural deduction
system.
Exerclse 4.5.2
PrOVe the fO110wing sequcnts within thc axionl systcnl prcscntcd in this scction:
1.卜 P‐→ P
2
(∼ P―
→P)― →P
3
p”P
∼P卜 Q
4
→ (Q→ R),Q卜 P→ R
c″ ス PrfR
CLASSICAL PROPOSIT10NAL
LOGiC:METATHEORY
5.1:NTRODuCT10N TO METALOGIC
Metalogic is the logical study of formal logical systems. In metalogic we study one
or more formal languages (e.g., the language of propositional logic) which, being
the obiects of our study, are called oblect languages. 'We must, of course, use
language to talk about an object language, but the language we use is usually not
the object language itself. We call it the metalanguage. Virtually everything said in
a logic textbook, except for the problems and exercises, is formulated in the meta-
language. Our metalanguage is English augmented with an assortment of variables
(e.g., Greek letters) and other technical devices.
Chapter 4 covered proofs formulated and carried out in an object language,
the language of propositional logic. The proofs we shall construct in this chapter
will be proofs about the object language, formulated and carried out in the meta-
language. The conclusions of these proofs are called metatheorems, and the proofs
themselves are called metaproofs. Reasoning in metaproofs often mirrors reason-
ing in the obfect language. Similar inference rules (e.g., modus ponens) may be
used, though usually they are used without comment, rather than being explicitly
cited and annotated. Over the years, logicians have developed a peculiar style and
rhetoric for expressing metatheorems. Part of what you will be learning here is
that style and rhetoric.
11:〕
114 CHnprrR 5
The letters 'QED' at the end stand for quod erat demonstrandum, aLatin phrase
meaning "which was to be proved." This is the logician's equivalent of a high five.
\u7hen you finish proving a meratheorem, writing these letters gives you
a little
rush.
Metatheorems differ greatly in form and content. Here is another simple
metatheorem that uses the valuation rules, rather than the formation rules:
This metatheorem simply puts into words what a truth table or tree would reveal.
But for that very reason it may be useful as an illustration of metatheoretic style.
Simple metatheorems may just be summaries or reminders of what we already
know; that is the case, too, with the next one.
But this next metatheorem is more general in scope. It combines the defini-
tion of a valuation, the formation rules, and the valuation rules to get the conclu-
sion that all formulas have exactly one of the values T or F on each of their
valuations:
PROOF:
■” ●一
〓 ■
一■●●
Though simple, this metatheorem is important for it confirms that our semantics
is bivalent-that is, classical.
Exercise 5.1
Prove the following metatheorems:
1. (Q (R v is a formula.
'(P
- " is true-S)))'
on the valuation in which'P'
2. '(P
- -P)' is false.
The conditions under which V(o & Y) = F are precisely those under which it is
not the case that "1"(o & Y) = T. (Check that the other rules imply that complex
formulas of other forms too are false iff they are not true. )
116 CHaprER 5
In formal logic, different kinds of proofs require different strategies. The same is
true in metalogic. Indeed, the same principles of strategy apply to both kinds of
reasoning. In both cases, strategy is governed mainly by the structure of the con-
clusion. For example, if the conclusion is a conditional statement (i.e., a statement
of the form @ * Y), then the best strategy is usually conditional proof. The object
language version of conditional proof was covered in Section 4.3, where it is called
conditional introduction (-I). Here we discuss conditional proof in the meta-
language.
In a conditional proof we suppose the antecedent @ for the sake of argument
and use it, perhaps together with oiher assumptions, to derive the consequent y.
The argument in which we derive Y is hypothetical, in rhe sense that it depends on
the supposition of the antecedent @, which we need not assert to be true. Because
of its hypothetical character, I like to think of this argumenr as a kind of fiction.
However, if we succeed in validly deriving the consequent y from the antecedent
@, then we certainly know this:
If we can do so, then we can discharge the supposition (A) and assert the desired
conditional statement (A) -t (C).
The problem now becomes how to derive (C) from (A). The argument is not
immediate, so we need some additional assumptions.2 As noted above, these are
likely to be definitions of major terms used in the metatheorem. Two major terms
used here are 'valid' and 'consistent'. The latter is used in two senses; statement
(A) concerns the consistency of a set of object language formulas, and statement
(C) concerns the consistency of a single object language formula, the conclusion.
So the next step is to look up the definitions of these terms. These definitions are
as follows:
(1) A sequent is valid iff there is no valuation on which its premises are true
and its conclusion is not true.
(2) A set of formulas is consistent iff there is at least one valuation in which
all members of the set are true.
(3) A single formula is consistent iff there is at least one valuation in which
it is true.
(These definitions are statements in the metalanguage concerning sequents or for-
mulas of the object language.)
Now the path of reasoning from (A) to (C) is easy to see. Since the sequent
in question is valid, by (1)there is no valuation on which its premises are true and
its conclusion is not true. But since the set of its premises is consistent, by (2) there
is at least one valuation on which these premises are all true. Hence on that valu-
ation (if no other) the sequent's conclusion is not untrue-that is, it is true. But,
then by definition (3), the sequent's conclusion is consistent-which is the conclu-
sion (C) that we were trying to prove.
That, of course, was just the hypothetical argument. Having completed the
hypothetical argument, we still need to discharge supposition (A) and assert our
conclusion (A) * (C). Here's what it looks like when we assemble the pieces:
of premises of a valid
Suppose thesetOrr...r@,of
premises of sorne valid conclusion Y is consis-
tent. Then (by the definition for a ls at
one valuation Y on which (Dr, . . ., 0,
of validity) there is no which
all true and Y is not true. Thus Y rs not untrue
'z By the way, be sure to keep straight which argument we are talking about
at which
time. We are talking about two arguments: a metatheoretical one that we are con-
structing and a formal one, technically a sequent, in the object language that the
metatheoretical argument is about. Metatheory is always working on two levels
like this, and that's one of the things that makes metatheory difficult.
118 Clapren 5
The indented part is the hypothetical argument, the "logical fiction."'We need not
know or care whether there is any such object language sequent as we are suppos-
ing here (though in this case, of course, there are many). The hypothetical argu-
ment takes us from the antecedent (A) to the consequent (C) of our conclusion.
Then the antecedent is discharged (indicated by ending the indentation) and the
conditional conclusion is asserted to complete the proof.
I have used parenthetical remarks to indicate where definitions are invoked
in the proof. In most metatheoretical writing, these remarks would be omitted; it
is assumed that the reader is sophisticated enough to realize that the argument is
by definition. But it helps when you are learning metatheory to remind yourself of
what you are doing by incorporating such remarks.
Notice how I used Greek letters as variables in the metalanguage. Such vari-
ables provide clear reference and help to condense the prose. I used Greek because
I didn't want the variables that belong to the metalanguage to be confused with
the P's, Q's, and so on that are part of the formal object language we are talking
about. If, for example, I had used the letter
P
instead of
Y
in the metatheorem, someone might have thought the conclusion was meant spe-
cifically to be the atomic formula 'P'. The metatheorem, however, does not specify
the form of the conclusion. To designate object language formulas without speci-
fying their identity, we need special variables in our metalanguage. And since we
don't want to confuse these metavariables-that is, variables of the meta-
language-with object language formulas, it's best to use a wholly distinct alpha-
bet. That's the reason for the Greek. When no such confusion could arise, however,
we will sometimes use the more familiar English letters as metavariables.
The purpose of the notation '{D,, . . . , @,' should be clear. This is just a way
of designating a list of some unspecified number (z) of formulas. We stipulate that
this notation allows the possibility that the list has no members so that z might be
0, unless otherwise specified. If n = 0, then by convention every valuation makes
@,, . . . , O, true, since there is nothing to make true. (The convention in question
is a convention that governs universally quantified statements-that is, statements
about all members ot euery member of a class-in classical logic generally. In
this case the particular statement at issue-'Every valuation makes @,, . . . , @,
11us'-15 a statement of the metalanguage. Further explanation of this convention
must await a full treatment of the semantics of quantifiers, which is given in
Section 7.2.)
Cressrcnr PRoposrrrolrerLocrc: Nrlrrarreonv 119
One final point about this metatheorem: Its hypothetical argument follows a
pattern that is very common in metalogical reasoning:
Then (by the definition of consistency for a set) there is at least one valua-
tion Y on which @,, . . . , @,are all true. But (by the definition of validity)
there is no valuation on which Or, . . . , Q,are all true and Y is not true.
The logical step here is simply the elimination of a double negation. The final
stage, repacking, puts the newly derived conclusion back into defined terms:
rue-that is, (by the definition of consistency) that the set {@r, . . . ,
Q,, -Yl is inconsistent.
Hence we have shown that if a sequent is valid, then the set consisting
of its premises and the negation of its conclusion is inconsistent.
Now suppose (again for conditional proof) that the set {O1,
. . . , @,, -Y) is inconsistent. This means that there is no valu-
ation on which @,, . . . , O,, and -Y are all true. Hence (by
valuation rule 1) there is no valuation on which Or,...,O,
are true and Y is not true, which is to say that the sequent O1,
...,@,FYisvalid.
Thus, if the set containing and the negation of its
conclusion is inconsistent, In summar5 we
have shown that a sequent its
premises and the negation of its QED
The proof consists of two conditional proofs, whose conclusions are assembled
into the biconditional at the end. Again, I am saying more here than the usual
sparse metalogical style permits. At most what is actually written when this kind
of proof appears in a journal article is the two hypothetical arguments (indicated
here by the indentations). The rest would be understood as implicit. Vith this
example it is also possible to combine the two proofs into a series of biconditional
inferences-something like this:
This style of proof is fairly common. The curly brackets '{' and '}' are the conven-
tional marks used to indicate the members of a set. I7e will employ this convention
from now on. For more on sets, see Section 7.1.
Exercise 5.2
Prove the following metatheorems by conditional proof:
The hypothetical argument begins with the supposition of the denial of the in-
tended conclusion. This is shown by the hypothetical argument to lead to the
contradiction that {O,, . . . , O,,} is both consistent and inconsistent. So the suppo-
sition is discharged (ending the indentation), and the desired conclusion is asserted
as proved.
Exercise 5.3
Prove the following metatheorems by reductio ad absurdum:
The metatheorems we have considered so far are extremely simple. More interest-
ing metatheorems use several different strategies, one nested inside another. For
example, in deriving the consequent from the antecedent in the hypothetical ar-
gument of a conditional proof, we might need to use a reductio strategy so that we
nest a reductio argument inside a conditional proof. Here is an example:
M[ETATHEOREM:If thc
conclusion of a second valid of
all the premises of both sequents
pRooF: Suppose for conditional proof conclusion of one valid
sequent is @ and the conclusion a second valid sequent
is -O.
Now suppose for reductio that the set consisting of all
the premises of both sequents is consistent. That is (by
the definition of consistency for sets), there is some val-
uation Y which makes each member of this set true.
CrassrcarPnoposrrtoxarLoctc: lVlerarHeonv 123
The trick in proving this metatheorem is to pay careful attention to the form of
the conclusion, that is, the metatheorem itself. The metatheorem is a conditional
whose antecedent is
(A) The conclusion of one valid sequent is @ and the conclusion of a sec-
ond valid sequent is -@.
(C) The set consisting of all the premises of both sequents is inconsistent.
So to prove this conditional, we suppose (A) for conditional proof (thus beginning
the proof with an indentation to indicate that we are engaged in a logical fiction)
and from (A) derive (C). Then we discharge (A) and assert the conditional conclu-
sion (the last statement of the proof).
But how can we derive (C) from (A)? The clue to follow here is that (C) is a
negative statement; it says that a certain set is inconsistent, that is, not consistent.
Negative conclusions are usually best proved by reductio. So within the hypothet-
ical argument of the conditional proof, we use a reductio strategy. We thus suppose
the denial of (C) for reductio. This is our second supposition, so we indent a second
time; we are now engaged in a "fiction within a fiction"-something like the play
performed inside Shakespeare's comedy A Midsummer Nigbt's Dream. I7e then
proceed from (A) by simple definition, as in the previous examples, and a contra-
diction follows quickly. This contradiction brings the "inner" fiction to an end.'We
discharge the reductio supposition and conclude that its negation is true. But this
negative conclusion is precisely the conclusion (C) that we were aiming for.
Exercise 5.4
Prove the following metarheorems by using reductio arguments inside conditional
proofs:
2. IftheconclusionYofasequent@,,...,@,, l-Yisvalid,thenthesequentitself
is valid.
3. If newpremises @,,,,,...,@,, are added to a valid sequent@r,...,O, l-Y,
thentheresultingsequent@,r. . . r@ur0,+r, . . ., O,,, F Yisvalid.
4. If O1, ...,@,, l- Y and Y, Y,, ...,Y, F @ are valid sequents, then @,,...,
Y-, Y,, . . ., Y, F O is a valid sequent.
5.5 MATHEMAITICAL]NDUCT:ON
R -R --P, ---R...
(the dots indicate that the series conrinues infinitely). Let us call this series S.
Now suppose that we want to prove that each item of S has the property of
being a formula. Of course, that's obvious from the formation rules, but we are
concerned with how to give a proper proof of it. The proof requires mathematical
induction. To prove by mathematical induction that each item of a series has a
given property F, we prove two things:
The proof of (1) is called the basis case of the induction; the proof of (2) is called
the inductive step. If we can prove these two things, then our work is done, for
together they logically imply the conclusion that every object in the series has the
property, even if the series is infinite. To see this, note that (2) is a universal state-
ment which implies each of the following instances:
If item t has F, then so does item 2.
If item 2 has E, then so does item 3.
If item 3 has E, then so does item 4.
...andsoon.
CrassrcarPRopostrtoruarLoctc: lVerarueonv 125
But (1) tells us that item t has F. So by modus ponens, together with the first
statement, item 2 has F. But then by modus ponens, together with the second
statement, item 3 has F, and so on. Thus by infinitely many steps of modus ponens
it follows that each item in the series has F. Of course we can't actually carry out
infinitely many steps of modus ponens. That's why we have the special principle
of mathematical induction. (\We wouldn't need it if all we ever had to worry about
were finite series.) This principle stipulates that if we have proved (1) and (2), we
can conclude straightaway that each item of the series has F; we needn't bother
with modus ponens. The validity of the principle is obvious.
In proofs by mathematical induction, the basis case is usually trivial. The
inductive step justifies the universally quantified conditional (2). The strategy is
always conditional proof. We suppose for conditional proof that some arbitrary
nthitemof the series has F (this supposition is called the inductive hypothesis) and
prove from this supposition that the (n + 1)st item has F as well. That proves the
conditional, and since the item considered was arbitrary, we can universally gen-
er alize the conditional.
Now in the problem we are considering, F is the property of being a formula.
'We
want to prove that every item of series S has this property. The basis case must
establish that'P'is a formula (which follows immediately from formation rule 1),
and the inductive step must show that if one item in the series is a formula, the
result of prefixing it with a negation sign is also a formula (which follows imme-
diately from formation rule 2). Hence the proof is easy. Here it is in proper meta-
theoretic form:
Basis Case The first item of S is 'P', which (by formation rule 1) is a
formula.
lnductive Step: Suppose that the is a formula. (This is the
inductive hypothesis; proof.)Now
the (n + 1)st item the nth with a nega-
tlon slgn. and the inductive
hypothesis) the {n
conditional ltcn■
so is the induction)
lten■ of s ls formula.
Again, this is more explicit than the usual metatheoretic style. In professional
writing, the labels 'Basis Case' and 'Inductive Step' and the parenthetical remarks
would be omitted, as would the last two sentences, which explicitly use conditional
proof and mathematical induction to draw the conclusions.
Mathematical induction enables us to prove that each item in a sequence has
a given property F. In the previous example, F was the property of being a formula.
126 CHaprrn 5
In the next example, F is the property of being logically equivalent to 'P', and the
sequence is:
M[ETATHEOREM:Each itcm Of
PR00F:
Basis Case: The first item of 7 is
truth value as 'P' on any valuation.
equivalent to 'P'.
lnductive Step: Suppose that the nth item tD of T is logically equivalent
to 'P'. That is, (D is true on any valuation on which 'P' is true
and false on any valuation on which 'P' is false. Now the (z +
1)st item is of the form (<D & P). On any valuation on which
'P' is true, therefore, both conjuncts of (O & P) are true; simi-
larly, on any valuation on which 'P' is false, both conjuncts of
(O & P) are false. Thus, by the valuation rule for conjunction,
(O 6a P) is true on any valuation on which 'P' is true and false
on any valuation on which (O & P) is false. Thus 'P'has the
same truth value as (O & P) on every valuation of both, and
SO(Φ &P),WhiCh iS the(κ +1)st
equivalent to 'P'
Thus (by conditional proof) it
cally equivalent tO `P' then
induction each ltein of T lS
zth item as its initial list contains exactly z paths. That is, F is the rather complex
property of being a disjunction whose tree contains tbe number of patbs desig-
nated by the numeral that subscripts its second disjunct. Despite the complexity of
this property, mathematical induction operates in precisely the same way as in the
previous metatheorem. Here is the proof:
The
formula is checked and 'P'with its subscript is atomic. And
Thus we
lt paths, then the
by using the (n + 1)st item of T as its initial list has
+ 1 paths. So (by mathematical induction) for all n, the tree
constructed by using the zth item of T AS its initial list has exactly n
QED
METATHEOREM: [CONCLUSION]
PR()(DF: se for reductio that [D OF usroNl
Nl
Hence (by reductio) [CO.NCLUSI oNl.
Basis Case:
varles IS
property F.
Inductive Step:
that the ath member of [SERIES]
Unpacking
Logical Manipulation
Repacking
(n + l)st F.
Hence (by conditional proof) we have shown that any n,
member of [SERIES] has property F, so does the (n + 1)st. Consequently
uslnS mathematica induction to com bine this conclusion with the
conclusion of the ba SlS casc all members of lsERrESl ha F.
CrassrcarPRoposrrroruarLocrc: lVerarneonv 129
Exercise 5.5
Prove the following metatheorems:
R -R --R ---R...
4. Each member of the following sequence of formulas
P,, (Pr.- P,), (Pr* (Pr- P,)), (Pr* (Pr.. (P.* P,))),...
is true in any valuation on which'P,'is true. (Hint: For each n, the (z + 1)st
member of the series is a conditional whose antecedent is 'P' subscripted by the
numeral for n + 1, and whose consequent is the zth member.)
5. If Y is a valid formula, then every member of the following sequence is a valid
formula:
Y, (YvP,), ((YvP,)vPr), (((YvP,)vPr)vP,),...
(Hint: Use mathematical induction inside a conditional proof.)
5.6 ALGORITHMS
27
82
+13
This initial symbol sequence is called the input to the algorithm. Then you perform
a series of precise, well-defined operations that yield a new sequence of symbols,
namely,
722
This is the output, or answer. The steps comprising the algorithm consist of adding
the individual digits (starting in the rightmost column), carrying the appropriate
numbers, and so on. This procedure, of course, works with any finite column of
numbers as input so that once you learn the algorithm, you can, at least in princi-
ple, add any column of numbers.
The symbols or characters that an algorithm operates on need not be numer-
als. The automatic "search and replace" operations available on most word pro-
cessors, for example, are simple algorithms that operate on the character set of a
computer (which includes the English alphabet), rather than just on numerals. Say
you want to replace all occurrences of the word 'Milton' in a document with the
word'shakespeare'. You put the cursor at the beginning of the document and invoke
the algorithm. The computer then runs through the entire document from beginning
to end, making the replacements you indicated. In this case, the input to the algo-
rithm consists of three symbol sequences: 'Milton', 'Shakespeare', and the initial
document. The output is the revised document, in which the word'Shakespeare'has
replaced the word 'Milton'. Here again the algorithm is a general procedure; it
operates not only on these three symbol sequences, but (in principle at least-
ignoring the memory limitations of computers) on any three sequences of letters.a
This is why mathematicians sometimes refer to algorithms as general procedures.
The concept of an algorithm carries with it some important presuppositions,
which are not always explicitly recognized. First, each algoritbm is defined only
ouer a prescribed character set, tbat is, a specified alphabet of symbols. Though
both sequences of numerals and strings of ordinary text are symbol sequences, you
can do addition only on sequences of numerals, not on strings of text. That is, the
generality of an algorithm is not absolute; it is limited by the kinds of symbols the
algorithm is designed to deal with. More specifically, each algorithm presupposes
a fixed character set upon which it works.
A character set is simply a finite set of discrete symbols. It may be as simple
as the binary alphabet of a computer (which has only two fundamental symbols,
often represented as 0 and 1 )s or as complex as the typographical system of English
a When the input to an algorithm consists of more than one sequence of symbols,
as it does here, these may be regarded as a single sequence in which the three
elements are listed in some conventional order (for example: term to be replaced,
term to replace it, document). We might in practice need additional symbols (such
as spaces, commas, or other special symbols) between successive members of the
sequence so that we can tell where one ends and the next begins. But by this
means, any finite set of sequences can be treated as a single sequence. Therefore
we lose no generality by thinking of an algorithm as operating always on a single
sequence of symbols.
5 Simpler still are character sets containing only one character; the abaci discussed
in Section 10.1 use in effect only a single character type: the counters that are
manipulated in their registers.
Classrcar PRopostrtonarLoctc: lVErarproRv 131
(which includes letters, both upper and lower case, numerals, punctuation marks,
etc.). Logicians, of course, are most interested in the character sets of logical lan-
guages. The character set for propositional logic, for example, consists of the
twenty-six capital letters of the English alphabet, the ten numerals 0-9 (for
subscripts), right and left parentheses, and the five characters for the logical oper-
ators. Among the most prominent algorithms applied to sequences of these sym-
bols (formulas or lists of formulas) are truth tables and trees.6
The character set presupposed by any pdrticular algorithm must be finite;
\We shall
that is, it must not contain an infinite number of fundamental symbols.
sometimes talk about infinite sequences of characters, but the character set itself
must be finite. (Thus these infinite sequences always contain repeated characters.)
This is a genuine limitation, though it might not seem so at first, for there are
symbol systems that can be interpreted as having infinitely many characters.
Consider, for example, the dial of a nondigital watch. If the hands move con-
tinuously-instead of in discrete jumps or ticks-then each configuration of the
hands might be thought of as a character representing a time. But between any
two distinct positions of a given hand, there is always an intermediate position so
that there are infinitely many of these "characters." Such characters (or sequences
of them) would not, therefore, be appropriate input for an algorithm. We'd have
to digitize them-that is, represent them in a symbol system with a finite character
set, before we could apply anything that could legitimately be called an algorithm.
The character set must not only be finite; its symbols must be distinct as well.
This does not preclude some variation. For example, all the following constitute
tokens of the letter 'U' in the character set of English:
UUUu,,UuUu
But what about this?
■′
′ FL
Is it a 'U', or an 'O' that didn't quite get closed at the top? In reality, there are
borderline cases-symbol tokens that could be classified either of two or more
ways. But the conception of an algorithm presupposes that such things don't hap-
pen, that each individual symbol is distinct and uniquely classifiable.
ln addition, inputs to and outputs from algoritbms must be sequences of
symbols. That is, they must be arrayed in a distinct linear order, like the text you
are now reading. By convention the sequential order of this text is from left to
6 Actually, truth tables and trees are not quite algorithms as we use them' since we
allow some choice as to which rule to apply next in a tree or which subformula
to analyze next in a truth table. Only if we adopted rules that rigidly determined
the order of these operations would trees and truth tables be algorithms, strictly
speaking. We could adopt such rules, just to make truth tables and trees conform
to the definition of an algorithm, but that would be more trouble than it would be
worth.
132 Cnapren 5
right and from the top to the bottom of the page, but other conventions could be
used, as they are in some languages. Below are some symbols from the English
character set that are not arranged in any clear sequential order:
B
R
x N
T
Q V
ZC
This sort of thing could not be input or output for an algorithm, unless we estab-
lished some convention that would impose a sequential order on it.
Moreouer, all input sequences dre presumed to be finite. lnfinite sequences,
like the sequence of numerals used for counting,
))+P(
for which the truth-table algorithm is not defined. In general, we designate those
symbol sequences for which an algorithm ls defined as its permissible symbol
sequences.
Noting the limitations of algorithms gives us a clearer conception of their
nature. An algorithm can apply only to finite, linear sequences of absolutely dis-
tinct symbols-and only to those that count as permissible sequerlces of the pre-
scribed character set.7
A final and crucial feature of algorithms is that they may be either terminat-
ing or nonterminating. A terminating algorithm is one that, given any permissible
input, will always yield its ourput after a finite number of steps. A nonterminating
algorithm is one that for at least one permissible input does not yield its entire
output after any finite number of steps. The procedure for adding columns of
numbers, for example, is a terminating algorithm. But the algorithm for counting
(using a given numeral, usually '1', as the starting point or input) is nonterminat-
ing. It's a fully definite computational procedure, but it never achieves completion.
Logicians are interested in algorithms because they want to know how much
of logic ian be reduced to mechanical computational procedures. Early in this
...rt,riy, some philosophers and mathematicians hoped that all the theses of logic
and mathematics could eventually be brought within the reach of terminating
algorithms-that the truth or falsity of any statement of logic or mathematics
co;ld be decided by finite calculations. They thought this could be accomplished
by completely formalizing logic and mathematics-that is, expressing them in
symbol iystems-and then devising the appropriate algorithms to operate on these
symbol systems. Accordingly, this line of research was called formalism. Formalists
hoped ro encode various fields of logic and mathematics in axioms (fundamental
assumptions) expressed in a logical language, and then to apply finite computa-
tional procedures (usually envisioned at the time as rules of inference) to these
axioms to determine the truth or falsity of any question expressible in the system.
But formalism failed.'We now know that it is impossible to answer alllogical
questions by finite computations, and we shall prove this when we consider the
undecidability of predicate logic. Nevertheless, for some restricted systems of logic,
the formalist dream can be realized. Propositional logic is one such system. For all
questions of validity (for both formulas and sequents), invalidity, consistency, and
so on, we have terminating algorithms (truth tables and trees) which give the
answers.
Here we are concerned with trees. It is obvious that the tree test is-or can
easily be transformed into-an algorithm (see footnote 6). What is not so obvious
is that it terminates for all inputs consisting of any finite list of sentences whatso-
ever. That requires proof. Our task in the next section is to construct a metatheo-
rem that proves this.
5.7 DECIDABILITY
8 Wecouldhaveshownthisforthetruth-tabletestaswell.Infact,fotthetruth-
table test it's obvious. We focus on trees' however, because the tree test can be
straightforwardly generalized to predicate logic; the truth-table test can't be.
134 CHAttER 5
(3) if a sequent is valid, the tree test classifies that sequent as valid (i.e., all
paths of its finished tree close).
lwe will prove proposition (1)in this section, proposition (2) in Section 5.8, and
proposition (3) in Section 5.9. Proposition (2) expresses the soundness of the tree
test; the test is sound in the sense that if it classifies a sequent as valid, the sequent
is in fact valid. Proposition (3) expresses the rest's completeness; the test is com-
plete in that it does not fail to classify any valid sequents as valid. uniting the
conditionals (2) and (3) into a single biconditional, we get a statement that ex-
presses the full accuracy of the tree test:
(a) The tree test classifies a sequenr as valid (i.e., all paths close) if and only
if that sequent ls valid.
Our first task is to prove that the tree test always terminates. The reason it
terminates is that whenever we apply a rule to a formula, each of the new formulas
produced by the rule is shorter (i.e., contains fewer characters) than the original
formula. Now formulas are not infinitely divisible; like material substances they
have smallest units, or atoms-namely, atomic formulas. Hence this shortening
process cannot go on forever. Eventually it has to stop.
This is the right idea, but it misses something: Although formulas grow
shorter and shorter within each path, the number of patbs increases each time we
apply a branching rule. In applying branching rules, might we not spawn so many
new paths that we creare more work than we complete and hence never finish? In
other words, even though no single path may ever grow infinitely long, might we
not generate so many new paths that the tree conrinues to grow-perhaps by
becoming "bushier and bushier"-without end?
In fact this cannot happen. But that is not so obvious; it requires proof.
our reasoning will fall into several parts, which we will express as lemmas
(short proofs preliminary to a major result). In lemma 1 we will show that because
of the shortening of formulas, each individual path in a tree must come to an end.
Then, to alleviate the concern that the tree might continue to grow forever anyway
(e.g., by multiplication of paths), we will prove in lemma 3 that in order ro grow
endlessly it would have to have an unending path, which lemma 1 will have shown
to be impossible. Before proving lemma 3, we shall prove another lemma, lemma
2, which provides a fact needed in proving lemma 3. Finally, we'll combine lemmas
1 and 3 into a metatheorem that proves the tree test always terminates. Before
embarking on our proofs, we need some definitions.
NoTE: For purposes of calculating tbe cbaracter count, tbe formation rules must
be followed strictly. This means tbat outer brackets may not be dropped; they are
included in tbe cownt.
Crassrcal PRopostrtoltaL Loctc: IVrrarltoRv 135
Lemma 1: If the character count of the tree's initial list is a, then each
path of the tree must be finished after at most z applications of the tree
rules to formulas on that path.
PRooF: Suppose the character count of the tree's initial list is a. Now
when any of the rules is applied to a formula on a path R each
of the resulting one-step extensions of P has a character count
at least one less than the character count of P (check this for
each of the rules).e Further, the minimum character count for
'q The biconditional rule, for example, allows us to check a formula of the form
(O * Y) and create one one-step extension to which we add (D and Y and another
to which we add -@ and -Y. The formula (O * Y) has three characters in addi-
tion to those in @ and Y, namely, '1','*', and')'. (Outer brackets are included in
the character count! ) But the first one-step extension omits all three of these char-
acters, keeping only O and Y, whereas the second adds only two characters in
addition to those in @ and Y, namely, two occurrences of '-'. Thus, since a
checked formula no longer counts, application of the biconditional rule reduces
the character count along the 6rst one-step extension by three and along the sec-
ond one-step extension by one. Similar reductions of the character count occur
with all the other rules.
10 The character count of a path drops to zero if the path closes; if it doesn't close,
the path must nevertheless be finished by the time its character count reaches
zero, since by that time all formulas are checked and so no further rules can be
apptied. Actually, the character count of an open path cannot drop as far as zero,
since some unchecked atomic formuIas or negations of atomic formulas remain
on the path.
135 CHaprun 5
Having shown that all paths must be finite, we must still prove that the tree
can't grow forever by endless proliferation of these finite paths. To make these
ideas precise, we add two more definitions.
That is, a path is infinitely prolongable iff the tree rules can be applied to make it
grow endlessly longer. This is just what lemma 1 rules out; that is, so long as the
initial list has a finite characer count, lemma 1 tells us that it cannot produce an
infinitely prolongable path. But we are still concerned about a tree growing end-
lessly in some other way-for example, by becoming infinitely ..bushy.,, The next
definition gives precision to this worry. It captures the idea of infinite growth in
general.
The infinite series Tr, Tr,... is a series of trees (or partial trees) generated by
applying rules starting with P, but not confining application of the rules to only
one path. Thus a path is nonterminating iff starting with that path we can apply
tree rules to formulas (perhaps among various branches into which that path
splits) forever. Nontermination is an apparently broader concept than infinite pro-
longability; a path might, it seems, be nonterminating by being able to grow end-
lessly more "bushy," as well as by growing endlessly longer. Actually, however, this
apparent difference is illusory, as we'll prove in lemma 3. But to prove lemma 3,
we first need to prove lemma 2:
P
●■ 一
Crnssrcar PRoposrtronaLLoctc: lVrrarnronv 137
The next lemma, which shows that a path can't grow infinitely in any sense
without being infinitely prolonged, is historically known as Konig's lemma:
'We
are ready at last to combine lemmas 1 and 3 into the major result of this
sectron:
pRooF: Suppose for reductio that this is not the case-that there is a
finite list I of formulas of propositional which a
tree that is not finished
Since sequents are always finite lists, and they remain finite when we negate
their conclusions, it follows that the tree test peformed on a sequent will always
6nish in a finite number of steps. The tree test is, in other words, a terminating
algorithm.
A test for validity is said to be sound if whenever that test classifies a sequent as
valid, it is in fact valid. In this section, we will show that the rree test is sound.I
We shall do this proof in two stages. First we will prove as a metatheorem that any
tree constructed from a consistent initial list has an open path. Then we will derive
t1 This is a different use of the term than when we speak of a sound argument-that
is, a valid argument with true premises. The soundness of the tree rules (or of the
rules of inference; see Section 5.10) implies nothing at all about rhe truth or fal-
sity of the premises. It is unfortunate that the same word 'sound' is used in both
of these ways, but the usage is so firmly established in the logical literature that
there is no point in bucking ir.
Crnssrcar PRoposrrroruaLLocrc: Nzlrratnronv 139
the soundness result explicitly as a corollary-that is, a result that follows easily
from something previously established.
To prove the conditional that if an initial list is consistent, then it always
yields an open path, we suppose for conditional proof that we have a consistent
'We
initial list. then unpack the notion of consistency as truth on some valuation,
which we shall call Y. The heart of the argument is to show that each time we
apply a tree rule, the resulting tree contains a path P whose formulas are all true
on V. But since all of P's formulas are true on Y, P cannot be closed because it
cannot contain both a formula and its negation (since these could not both be true
on Y). Hence P must be an open path.
That's the proof in a nutshell. Here it is in greater detail:
on "l/, then (by valuation rule 1) that path cannot contain both
a formula and its negation, and hence must be open. To prove
that each tree in the series contains a path all formulas of
which are true on V, we use mathematical induction:
Basis Casq The first member of the series is 7,, which is L
itself, and by hypothesis each member of L is true on Y.12
Inductive Step: Suppose (inductive hypothesis) that the zth
item 7,, of the series (where n < zl contains a path P all
of whose formulas are true on V. Now the (z + 1)st
item, T, * ,, is formed by a single application of a rule
12 When I say that this is true "by hypothesis," I refer to the fact that we are operat-
ing under the supposition (for conditional proof) that L is consistent, which
means there must be a valuation on which all of its members are true. We have
labeled that valuation "7". Thus each member of L is true on Y.
140 Craprrn 5
true
an open path.
Hence we have shown (by conditional proof) that if L is consistent,
then there is an open path through any (finished or unfinished) tree T
obtainable from L by the tree rules. QED.
'We
now use this metatheorem to prove that the tree test is sound. A test for
validity is sound, once again, if whenever that test classifies a sequent as valid, that
sequent is in fact valid. What it means for the tree test to classify a sequent as valid
is that, given the premises and the negation of the conclusion as an initial list, we
13 Take, for example, the disjunction rule v. Suppose this is applied to a formula
@ v Y which occurs on a path all of whose formulas are true on some valuation
"l/. Then, since @ v Y itself is true on Y, by the valuation rule for disiunction
either @ or Y must be true on V. Therefore, since the rule v produces two one-
step extensions of the path, one of which appends @ to it and the other of which
appends Y, all the formulas of at least one of these one-step extensions of the
path must be true on Y.
Crassrcnr PRoposrrronarLocrc: IVetarrroRv 141
get a tree all of whose paths close. The argument from our metatheorem to sound-
ness is relatively simple:
pRoor: Suppose (for conditional proof) that the tree test classifies a
the
tree close.
lS
lt lS In ねct
A radical skeptic might wonder what this soundness proof really proves.
If, for example, someone had doubts about the validity of sequents proved by
those doubts would hardly be allayed by a metatheorem established by con-
-I,
ditional proof-that is, the same pattern of reasoning at a different linguistic level.
As a response to such a person, the metaproof would be in some sense circular-
assuming the validity of one of the very patterns whose validity it purported
to prove.
Metatheoretic proofs are not, however, intended as responses to radical skep-
tics. To a person who doggedly doubts the elementary rules of logic, there is no
effective logical response. Rather, the point of this soundness proof is to show,
given a prior understanding of logic, that any sequent judged valid by the tree test
is valid on classical semantics-the semantics described in Chapter 3. The sound-
ness proof provides not a wholesale assurance that classical reasoning is irrefut-
able, but the more modest assurance that the tree rules validate only sequents they
ought to validate, given classical semantics. It presupposes, moreover, a willingness
to use what are in effect the rules of classical logic in the metalanguage. But what
alternative is there? If we did not grant the validity of some sort of reasoning
somewhere, we would never accept any conclusion and never come to systematic
insight about anything. That is the radical skeptic's game, but it is a game that
precludes much intellectual adventure.
Exercise 5.8.1
In the proof of the main metatheorem of this section, it is necessary to verify for
each tree rule other than - that when applied to a path whose formulas are all true
142 Crnprrn 5
on some valuation V, it yields at least one one-step extension of that path whose
formulas are all true on V or on some expansion of y. I did this for the v rule in
footnote 13. Write out the necessary verifications for the rules --, &, -&, -V, -,
-+, e, and -*.
Exercise 5.8.2
Using the metatheorem of this section as an assumption, prove the following
corollaries:
1. If the tree test classifies a formula as inconsistent, it is in fact inconsistent.
2. lf the tree test classifies a formula as valid, it is in fact valid.
3. lf a formula is contingent, the tree test classifies it as contingent.
In this section we prove the completeness of the tree test. A test for validity of
sequents is complete iff it classifies all the valid sequents as valid; in other words,
if a sequent ls valid, the test classifies that sequent as valid. To prove that the tree
test is complete, we proceed as we did in proving that it is sound. That is, we first
prove a general metatheorem and then append a simple corollary that proves the
completeness result.
The general metatheorem is expressed as the following conditional:
If there is an open path through a finished tree, then its initial list is
consistent.
To prove this conditional, the first step is to suppose for conditional proof that
there is an open path P through a finished tree. 'We then show how to construct a
valuation V on which all the members of P are true. But given that all members of
P-including the initial list itself-are true on V, by the definition of consistency
the initial list is consistent.
In showing how to construct'l/, we first define the notion of formula length.
The length of a formula is the number of characters it contains, excluding sub-
scripts. Thus, for example, the length of '-(P & Q)' is 6 and the length of '--P12'
is 3. Here is the proof in full regalia:
.PR00F: Suppose
through a finished tree. Since this tree is
CrassrcnrPRoposrrrotarLocrc: [\zlrraruronv 143
ta That is, only atomic formulas or negations of atomic formulas remain unchecked.
This must be the case, since otherwisc more rules could be applied on P, which is
open, and so the tree would not be finished.
ls Recall that formulas count as occurring on a path only if they are listed there as
whole formulas, not if they are merely parts of other formulas.
The point of making all sentencc letters that do not occur on P false is that
some sentence letters in the initial hst may not occur either negated r>r unnegated
along the path. This signifies that, given the other truth values determined by the
path, their truth value does not affect the rruth of formulas of the initial list. For
the sake of definiteness, we de6ne "ll in a way that makes those sentence letters
false, but this choice is arbitrary. Our definition also, ofcourse, makes senrence
letters whose negations occur on P false, and this is not arbitrary.
r6 Formulas of length 1 (sentence letters) are true because they are assigned thc
value 7 by V directly. And formulas of length 2 (negated senrence letters) are true
because, since their sentence letters do occur on path ( these sentence letters have
been assigned the value I bv'ii.
r7 'We know this because a finished path by definition contains no unchecked sen-
tences of length 3 or m<>re. Any checked sentence is a sentence to which a rule has
been applied.
1M CHnprun 5
Exercise 5.9
Using the metatheorem of this section as an assumption, prove the following
corollaries:
l. lf a formula is inconsistent, then the tree test classifies it as inconsistent.
2. lf a formula is valid, then the tree test classifies it as valid.
3. If the tree test classifies a formula as contingent, then it is contingent.
18 If, for example, the rule & is applied to a coniunction @ & Y, it will produce the
formulas o and Y on each path below it. But by the valuation rule for coniunc-
tion, if these two formulas are true, then @ & Y itself must be true. Similarly, if v
is applied to a disjunction (D v Y, each new path it produces will contain either @
or Y. But by the valuation rule for disjunction, if either of these two formulas is
true, then @ v Y is true. Thus, if either of these rules yields only true formulas
along a given path, then the formula to which it is applied must be true. The same
result holds for the other rules.
CLassrcarPnoposrrroNeLLocrc: [VlrrarHeony 145
A
2
1 8cE
3
l&E
4
2,3 8ζ I
5
48こ E
6
48`E
P
7
5,6 8cI
8
7 8ζE
9
oヽ
78こ E
・
This pattern can clearly be iterated ad infinitum. Nobody would do this in practice,
of course, but it is a common experience in working with inference rules to reason
in circles, repeatedly deriving what you have already proved or assumed. This
simple example shows that in principle you could do so forever.
So inference rules by themselves are not a decision procedure in the way
truth tables or trees are. It is, however, possible to design a terminating algorithm
for generating proofs by rigorously specifying the order of application ofthe infer-
ence rules. One way to do this is to make proofs mimic trees. If a sequent is valid,
we know from the completeness of the tree test that all the paths of the tree for
that sequent close. The tree procedure is closely akin to a reductio (-I) proof in
which the negation of the conclusion is shown to lead to contradiction. If we had
a terminating algorithm for reliably converting trees for valid sequents into such
proofs, then by using it we could avoid the kind of infinite regress illustrated
le I put the word'proof in quotation marks because an infinite structure like this is
not a genuine proof. A proof always has a final line on which its conclusion is
displayed.
146 Cuaptn 5
above. The whole procedure would, however, be parasitic upon the tree test. That
is, we would need to construct a tree first to determine whether or not the sequent
was valid; then, if all paths closed, we would convert that tree into a proof. For
invalid sequents, whose paths do not all close, prior performance of the tree test
would prevent us from attempting a proof. Thus an algorithm for converting trees
of valid sequents to proofs would have little value in itself, since once we have
determined the validity of a sequent by the tree test, it is redundant to construct
a proof.
Yet such an algorithm would have at least one valuable implication: It would
demonstrate the completeness of the ten basic rules. For if we could convert any
tree for a valid sequent into a proof of that sequent, then the following would
be true:
( 1 ) If the tree test classifies a sequent as valid, then that sequent can be
proved using only the ten basic inference rules.
Putting this together with the completeness of the tree test (which we proved in the
previous section):
(2) If a sequent is valid, the tree test classifies that sequent as valid,
we obtain the conclusion:
(3) If a sequent is valid, then that sequent can be proved using only the ten
basic inference rules.
But (3) asserts that the rules are complete. Since we have already proved (2),to
prove the completeness of the inference rules, then, we need only prove (1)-that
is, to show how to convert a tree for a valid sequent (a tree all of whose paths
close) into a proof that uses only the ten rules. Before defining a general method
for doing this, let's do the conversion for some specific examples.
Our first example is the valid sequent'P & Q F P v R'. Our aim is to show
how to convert the tree for this sequent into a proof.'We shall do this by construct-
ing the tree and the corresponding proof side by side. The first step of the tree test
is tq write the initial list, consisting of the premises and the negation of the conclu-
sion. The tree test reveals the inconsistency of this initial list (assuming the sequent
it represents is valid) by showing that each possible way in which all of its formulas
might be true leads to contradiction.'We may think of this as a reductio strategy in
which the negation of the conclusion is hypothesized in order to show that, given
the premises, it leads to absurdity. Thus we can begin to construct a proof that
mimics the tree test by assuming the sequent's premises and hypothesizing the
negation of its conclusion for indirect proof. Thus, for the sequent 'P & Q F P v
R', the tree and the corresponding proof begin as follows:
cation is unimportant. Let's begin by analyzing 'P 6a Q'. The corresponding move
in the proof is to apply &E twice:20
P ︿
P Q
3
3
18こ 18こ E
Ч
4
4
18こ 18こ E
The occurrence of 'P 6a Q' in the tree should now be checked off. The next step
in constructing the tree is to check '-(P t R)' and analyze it into '-P' and '-R'
using the -v rule. In the proof these same formulas may be deduced by convert-
ing '-(P v R)' into '-P & -R' by De Morgan's law (DM) and then using two steps
of &E:
P R
5
5
∼P8こ ∼R 2:[)ヽ4
∨ V
” ”
一 ”
6
6
∼P 5&こ E
7
∼R 5遅 にE
'We saw in Section
DM, of course, is a derived rule, not one of the ten basic rules.
4.4, however, that derived rules are merely abbreviatory devices; anything proved
with derived rules can also be proved with the ten basic rules so that the use of
derived rules here is legitimate.
Both 'P' and '-P' now appear on the tree's one path' The next step in the
tree, therefore, is to close this path using the negation rule. We shall think of the
'X' that closes the path as representing a contradictory formula-specifically,
'P 6a -P'. The proof also contains both 'P' and '-P'; thus the corresponding move
in the proof is to derive 'P & -P' by &I:
7.X 3,5 8。 │
P8(∼ P 3,68こ I
The tree, which represenrs only the derivation of a contradiction from the premises
and the hypothesis of the negated conclusion, is now complete. But to finish the
proof we need ro end the hypothetical derivation and deduce the conclusion. This
takes two more steps, one of -I and one of -E:
9. --(P v R) 1-8 -I
10. PvR 9-F.
We have now converted the tree for the sequent 'P & Q l- P v R' into a proof of
that sequent.
TLis example was unusually simple, since the tree did not branch. Irhen
trees branch, the corresponding proofs involve uses of vE within the overall -I
strategy, and things become more complicated. Take, for example, the valid se-
quent 'P v Q, -P F Q', which expresses one version of disiunctive syllogism. As
before, we begin the tree by listing the premises and negation of the conclusion,
and we begin the corresponding proof by assuming the premises and hypothesizing
the negation of the conclusion:
20 Only the first application of &E, the one at line 3, is essential to the proof, but we
are concerned here with proofs that mimic trees, not with proofs that are maxi-
mally compact. The extra steps are harmless'
148 Craprrn 5
l
I
D ■ H(br→ I)
5
I P8こ ∼P 2,48こ I
6
Q lv Q H(br→ I)
8
D
Q→ (P&∼ P) 7-8-→ I
0
P&こ ∼P l,6,9∨ E
The two conditional proofs constructed in preparation for vE represent the two
branches of the tree. The hypothetical derivation at lines 4-5 of the proof repre-
sents the left branch of the tree. The 'X' on the left branch of the tree corresponds
to the contradiction 'P & -P' at line 5 of the proof. The hypothetical derivation at
lines 7-8 represents the right branch of the tree, and the contradiction 'p & -p, in
the proof (line 8) represents 'X'that ends the right branch of the tree. But here
'P & -P' is obtained, not by &I from 'P' and '-P', but by the derived rule EFQ (ex
falso quodlibet; see Section 4.4) from 'Q' and '-Q'.
This use of EFQ is important. Though different formulas (such as ,p' and
'-P' or 'Q' and '-Q') -ay lead us to close paths in the tree, each occurrence of ,X,
must be represented by the same contradictory formula in the proof. This is be-
cause each of the two conditionals used in vE (here the conditionals appearing at
lines 6 and 9) must have the same consequenr. So to apply vE (which we do here
at line 10) we need to derive the same contradiction from each hypothesis. EFQ
will always enable us to do this. (EFQ is, of course, not one of the ten basic rules,
but we saw above that use of derived rules here is legitimate. ) ln fact, to standard-
ize our procedure, we shall arbitrarily stipulate tbat'X' in any tree altuays repre-
sents the formula'P dt -P' in the proof.
The conditionals on lines 5 and 9 of the proof and the conclusion derived by
vE at line 10 do not correspond to any particular formulas in the tree. They are,
rather, part of the apparatus of disjunction elimination, which ensures that contra-
dictions derived along different branches of the tree also follow from the formula
from which those branches stem.
CrassrcarPRoposrrroruarLocrc: MrrnrrroRy 149
As in the previous example, the tree represents only the derivation of a con-
tradiction from the hypothesized negation of the sequent's conclusion. To com-
plete the proof, we must end this hypothetical derivation and apply final steps of
-I and -E:
Indeed, any proof derived from a tree by the method illustrated here must end in
this way.
The full algorithm for converting trees for valid sequents into proofs may be
stated as follows:
Table 5.1, as noted in step 2, provides instructions for converting each application
of a tree rule into a series of steps in the proof. It remains only to verify that this
algorithm performs as advertised.
TABLE 5.1
instructions for Converting¬ rees for Vlalid Sequents into Proofs
(continued)
150 Cuapren 5
TABLE 5.1
lnstructions for Converting Trees for Valid Sequents into Proofs (continued)
Disjunction (v) If an open path Hypothesize @, aiming to derive 'P & -P'
contains an unchecked formula and then deduce O (P & -P) by Next,
- -I.
of the form ((D v Y), check it and hypothesize Y, again derive'P & -P', and then
split the bottom of each open path obtain Y - (P & -P) by --1. Finally, use vE to
containing this newly checked deduce 'P & -P' from (O v Y), (D (P & -P),
-
formula into two branches; at the and Y * (P & -P). This procedure always
end of the 6rst write @, and at the works if the sequent being tested on the tree
end of the second write Y. is valid, since in that case all the paths below
(O v Y) must close, and each closed path is
converted into a derivation of'P & -P' in the
proof (see negation rule). If there are further
applications of branching rules below (O v Y),
these will also be converted into derivations
of'P & -P' by further applications of the
procedure for disjunction.
(continued)
CrasstcarPRoposrtroruarLocrc: [VrrarHeoRy 151
TABLE 5.1
lnstructions for Converting Trees for Valid Sequents into Proofs (continued)
a
check it and list O above -Y at ∼(Φ → Ψ )
lD C
the bottom of every open path ∼Φ ∨Ψ H(br∼ I)
月 u C rl 00 ln ・1
formula. (Φ→ Ψ)&∼ (Φ→Ψ) a,C&I
―(∼ Φ ∨Ψ ) b― d∼ I
∼∼Φ &│∼ Ψ eDM
―∼Φ f8`E
Φ g∼ E
∼Ψ f8cE
(Here @ is proved at line h and -Y at line i.)
c― g∼
h
∼Φ I
i
∼Ψ a,h← →MiT
l
∼Φ &∼ Ψ h,i&I
k
(∼ Φ &∼ Ψ))b,k&I
∼((Φ &Ψ )∨
m∼ ∼((Φ 8(Ψ )∨ (∼ Φ &∼ Ψ)) b-1∼ I
n(Φ &Ψ )∨ (∼ Φ &∼ Ψ) m∼ E
・
「
hCn apply thC direCtiOnS fOr diSiunCtiOn tO
(Φ &Ψ )V(∼ Φ &∼ Ψ ).
TABLE 5.1
lnstructions for Converting Trees for Valid Sequents into Proofs (continuedl
k|
I
@*Y e.i-l
I | (<D*Y)&-1o*Y1 a,k&l
m --((o & -Y)v (-o & Y)) b-l -I
n & -Y)v (-O & Y)
(<D m -E
Then apply the directions for disjunction to
(o & -Y)v (-<D & Y).
2r In the case of the two rules for the biconditional, * and *, which produce
-
branches beginning with two formulas each, the hypothesis corresponding to
each branch in the proof is a single formula-the coniunction of these two
formulas.
Crassrcar PRoposrronarLocrc: lVlerarsEoRv 153
(by vE) from the formula from which the subbranches origi-
nate, Thus, whether or pOrtion of
the pFoOf derlvatlonll‐ ‐
. . . , @, and the hypoth-
Therefore we may apply step 3
QED
Finally, we shall show that the system consisting of the ten basic inference
rules is sound-that is, that any sequent provable by these rules is valid. \7e have
already seen in Sections 4.2 and4.3 that each rule individually is valid-that is,
that there is no counterexample to any instance of any of these rules. To prove a
'We
sequent, however, we apply these rules successively. must, then, show that
invalidity does not somehow creep into a proof as a result of this succession. In
order to show this, it will be useful to define the notion of a corresponding sequent
to a line of a proof.
The corresponding sequent for a given line is in effect what is proved at that line.
To illustrate, consider this proof of 'P*Q, -Q F -P'. Corresponding sequents are
listed to the right.
154 CHnpren 5
↓Q P
P 一 ︱
1
Q A P→ Q卜 P→ Q
2
A P→ Q,∼ Q卜 ∼Q
3
2,3 8cI
6
P→ Q,∼ Q卜 ∼P
”
3-5∼ I
Since all hypothetical derivations must end before a proof is finished, the
corresponding sequent for the last line of any proof is just the sequent whose
premises are the proof's assumptions and whose conclusion is the proof's
conclusion-that is, the sequent to be proved. Thus, if we can show that the
corresponding sequent for any line of any proof is valid, it will follow that the
corresponding sequent for the last line of any proof is valid, and hence that any
sequent provable by the ten basic inference rules is valid. ActuallS since inference
rules may apply to any earlier lines, it is easier to prove something apparently a
little stronger than this-namely, that each line and all lines preceding it corre-
spond to valid sequents. This can be done by mathematical induction on the num-
ber of lines in the proof. The induction appeals frequently to the following lemma,
which was problem 4 of Exercise 5.4:
METATHEOREM: Let P be any proof using only the ten basic infer-
ence rules; then for each line of P, the corresponding
lines up to and including that line are valid.
PRooF: The lines of P form a series,
ical induction. The property which we show t0 each
are valid.
22 Notice that none of the original hypotheses or assumptions @,, . . . , !D- can be
dropped, since that would indicate that --O appears in a hypothetical derivation
that has ended, so that neither -E nor any other rule may be applied to it.
156 CuaprrR 5
Suppose that
-E is applied to two lines each of whose
corresponding sequents is ualid. The formulas on these
lines are therefore of the forms @ * Y and @, and tbeir
corresponding sequents haue tbe forms @r,. . . , O- t- O
* Y and Ar,. .. r4,, F O, where m> 0 and n> 0. The
line obtained by the application of
-E is of the form Y,
and its corresponding sequent has the form Y,, . . . , Y,
l- Y, where p ) 0 and @r,...,@* and A,r... rL,are
included among Y,, . . . ,Yr." Now we saw in Section
4.2 that O * Y, O l- Y is a valid form. Hence, since
O,,. . .,@-l @ * Y is valid, it follows by the lemma
that @,, . . . ,@^, O F Y is valid. Given this and the fact
that A,, . . . , L, F tD is valid, it follows again by the
lemma that O,, ... r@,nrA,,...,& F Y is valid. But
srnce O,, . . , , @,, and A.,, . . . , A, are included among
Y,, . . . , Y, from this by problem 3 of Exercise 5.4 we
may infer that Y,, . . . , Y, F Y is valid. But this is the
corresponding sequent for the line obtained by *E.
Therefore, if *E is applied to a pair of lines, each of whose
corresponding sequents is valid, the result is a line whose cor-
responding sequent is valid. Finally, we show that if -I is ap-
plied to a series of lines all of whose corresponding sequents
are ualid, the resub is a line whose corresponding sequent is
u4lid.
Suppose that -I is applied to a series of lines all of whose
corresponding sequents are valid. For -I to be applicable,
the first such line must contain a hypothesized formula
@ and the last must contain a contradiction Y & -Y that
is derived from @. The corresponding sequent of this last
line must therefore have the form O, O,,- . . . , @*F Y &
-Y, where m > 0. Since we have supposed this sequent
to be valid, there is no valuation on which O, 6,, . . . ,
tD- are all true and Y 6c -Y is untrue. But since Y & -Y
is untrue on all valuations, there is no valuation on
which <D, Or,... ,Q^are all true. Hence by valuation
rule 1, there is no valuation on which O,, . . . , O- are all
true and -@ is not true. Therefore the sequent Qr, . , . ,
O,,, F -O is valid. But since application of -I ends the
hypothetical derivation from (D, leaving only (0,, . . . ,
O* as the assumptions or hypotheses whose derivations
have not ended, this sequent is just the corresponding
sequent for the line obtained by -I.
From this result, the soundness of the ten rules follows as a corollary:
must have ended before the last line, the corresponding se-
quent for this last line is just O,, . . . , O, F Y (see definition of
corresponding sequent). But by the previous metatheorem, the
corresponding sequent for any line of any proof is valid. Hence
@r, . . ., (D, l- Y is valid.
Hence, if a sequent can be proved using only the ten basic inference
rules, then that sequent is valid. a ED
Having shown that the ten basic inference rules are sound and complete (i.e.,
that they enable us to prove a sequent of propositional logic iff it is valid), we now
know that they enable us to prove exactly the sequents we should be able to prove,
given the classical notion of validity.
Exercise 5.10.1
Prove for each of the rules &I, &E, vI, vE, -E, *I, and *E that if applied to
lines whose corresponding sequents are valid, they yield a conclusion whose cor-
responding sequent is valid.
158 Craprun 5
Exercise 5.10.2
1. A set of rules is consistent iff there is no formula @ such that both <D and -<D
are provable as theorems from these rules. Use the soundness of the ten basic
inference rules to prove that these rules are consistent.
2. Use the soundness and completeness of the ten basic inference rules to prove
that a formula of propositional logic is valid (tautologous) iff it is a theorem.
」
中ァE「
●▼
│■ ■││■‐││■ │
‐ ‐
. ││‐
.│ ││... . . .
│‐ │. .
C月 ハ ρ ア ど R
CL`ASSICAL PREDICAITE
LOGIC:SYNTAX
Propositional logic is the study of how validity and related properties arise from
formal configurations of the operators '-'r'&' ,'v' ,'-' , and '*'. But these are not
the only logical operators. Other expressions also contribute to validity. Consider,
for example, the argument
All women are mortal.
Cleopatra is a woman.
Cleopatra is mortal.
This is clearly valid in the informal sense described in Chapter 1. But what makes
it valid? None of the five propositional operators are seen to occur here so that if
we were to attempt a symbolization in propositional logic, the best we could do
would be something like 'P, Q F R'. But this sequent represents any argument with
two premises, and it is invalid. Plainly there is some syntactic and semantic struc-
ture within these sentences that accounts for the argument's validity, but iust as
plainly propositional logic does not enable us to represent it.
The atoms of propositional logic are sentence letters. But, as in the preceding
argument, the sentences these letters represent have internal components-sub-
jects, predicates, and modifiers of various sorts-whose arrangement may affect
161
162 Cuaprrn 6
t This curious backwardness is rhe fault of Gorrlob Frege, who invented predicate
logic late in the nineteenth century. Frege thought of predicates as designating
functions which produce truth values when applied to objects. Function symbols
in mathematics are usually written before the names of objects to which they ap-
ply, and Frege adopted this mathematical convention.
Cussrcar Pnrorcnre Loctc: Svnrax 163
Vx(\7x * Mr)
This is a formula of predicate logic. Unlike the other expressions we wrote
to produce it, this formula is not a sentence, but (like the formulas of propositional
logic) a representation of a sentence form. By itself, it has no specific meaning,
though when we assign meanings to the predicate letters (as we have done) it does
mean something-in this case, "Al1 women are mortal."
The remaining two sentences of our argument both contain the name'Cleo-
'We
patra'. shall use lowercase letters 'a' through 't' as names for people or things,
so 'c' is the obvious choice for Cleopatra. Since, as we noted above, subjects
customarily are written after predicates, 'Cleopatra is a woman' becomes ''$7c'and
'Cleopatra is mortal' is 'Mc'. Thus the entire argument may be formalized as the
sequent 'Vr(\Jfr - Mx), !7c F Mc'.
may validly deduce ''W'c *
Now it is obvious that from 'Vx(\i7x - Mx)'we
Mc'-in English:
-VxMx
This is a perfectly intelligible formula, but what it says is "it is not the case that for
all x, x is mortal"-or, more compactly, "not everything is mortal." This state-
ment, unlike 'Nothing is mortal', is compatible with the existence of some mortal
things.
Consider now the sentence
That is,
Vr(Wx * -Mr)
'No' in this context is thus analyzed into universal quantification over a condi-
tional whose consequent is negated.
In addition to 'V', which means "for any" or "for all," predicate logic con-
tains a second quantifier, '3', which means "for at least one." The English word it
is most commonly used to symbolize is 'some'. consider, for example, the sentence
Using 'F' for the predicate 'is a father' and 'G' for 'is a gorilla', we may symbolize
this sentence as
lx(Fx & Gx)
which is an existentially quantified conjunction. Actually, in English, the word
'some' tends to mean "at least two," so there may be some slippage of meaning in
the formalization. This slippage is usually not too troublesome, but it should not
be forgotten. The English expressions 'there is' and 'there exists' are perhaps better
translations of 'l'than 'some' is. Hence the preceding formula can also be read as
There is an x such that r is a father and x is a gorilla.
or
There exists an x such that x is a father and x is a gorilla.
Accordingln 'l' is called the existential quantifier.
\We might also wanr to say (because it is
true) that some fathers aren't goril-
las. This means
Notice that it is wrong to render the true statement 'some fathers aren't go-
rillas' as
which is to say that no fathers are gorillas-which, of course, is false.3 But this
itself tells us something interesting: Statements beginning with 'no'can be rendered
into predicate logic in either of two equivalent ways. For example, we can sym-
bolizi the statement 'No women are mortal' either with the universal quantifier as
Vx(Wx - -Mx)
as we did earlier, or with the existential quantifier as
That is, "There exists an r such that x is a unicorn and x has one horn." This, of
course, is false, since unicorns don't exist. Suppose, however, that we incotrectly
formulated this same statement using the conditional operator instead of the con-
junction, that is,
Now what we have said is something true bvt uery strange: There exists something
such that if it is a unicorn, then it has one horn. We are not saying that anything is
a unicorn, nor are we saying that anything has a horn. (The statement we were
trying ro formalize said both of these things.) The sort of object described by this
erroneous formulation might be anything at all. It is true of my neighbor's cat, or
the state of Alaska, or even the number 47 that if it is a unicorn, then it has one
horn.
3 At least as of this writing. Gorillas being endangered, it may soon to our sorrow
become true.
166 CHaprrn 6
Some English predicates, such as'loathes', are transitive, taking both a sub-
ject and an object. These need two names to make a complete sentence and hence
are called two-place predicates. The usual convention is to write them in the order
predicate-subject-object. So, for example, the statement 'Beth loathes Carl' may
be formalized (using 'b' for 'Beth', 'c' for'Carl', and 'I- for loathes) as'Lbc'. Two-
place predicates may be combined with quantifiers and the propositional operators
for still greater variety of expression. Here are some examples:
There are things that loathe themselves and things that don't.
lxl-xx &-lx-Lxx
158 CHaptn 6
Here the first three occurrences of 'x', being governed by the first existential quan-
tifier, refer to the things that loathe themselves, and the second three occurrences,
being governed by the second existential quantifier, refer to the things that don't.
Thus two occurrences of the same variable must refer to the same thing only if
they are governed by the same quantifier. Some people find this confusing and
would prefer to formalize this conjunction as
lxLxx k ly-Lyy
That's fine. But so is the first way we did it.
It is not fine, however, to have two quantifiers governing the same variable,
as in 'vrvrlxx' .Here we can't tell which variable is governed by which quantifier.
I7e shall not even count such expressions as formulas.
I07hen universal and existential quantifiers occur in the same sentence, the
sentence is usually ambiguous. Consider the sentence 'something loathes every-
thing'. This has the following two meanings:
1. There is some one being (think of Danre's Lucifer, for example) that
loathes everything, or
2. Everything is loathed, but not necessarily by the same being in each
case. (Thus Beth may be loathed by Carl, Carl by himself, and so on,
though perhaps no one being loathes everything.)
Because the founders of predicate logic were trying to create an ideal language,
formulas of predicate logic are, by design, perfectly unambiguous. consequently
when we formalize 'Something loathes everything', we must decide on one or the
other of these meanings. The difference between them is reflected in the formalism
by the ordering of the quantifiers. Converting each reading into semilogical Eng-
lish, we obtain
1. lxYyLxy
2. YylxLxy
By forcing us to choose one meaning or the other, predicate logic sensitizes us to
ambiguities of natural language that might orherwise go unnoticed.
Ambiguiry that arises from mixing universal and existential quantifiers in
English is called a quantifier scope ambiguity. Because English grammar allows
much greater flexibility in the placement of quantifiers within a sentence than do
the formation rules of predicate logic (which we shall consider in the next section),
it has no reliable way of indicating their scope (the part of the sentence they
govern). But, as we shall see, the scope of a quantifier in a formula of predicate
logic is always definite and clear.
The order of the quantifiers matters, however, only when universal and ex-
istential quantifiers are mixed. There is no difference in meaning, for example,
CrassrcnL Pneotcaru Loctc: Svrurax 159
between 'YxYyLxy' and 'VyVxLxy'. They both express the thesis of universal
loathing: Everything Ioathes everything.
Two-place predicates do not always represent transitive verbs. Some repre-
sent predicate phrases of other sorts. The phrase 'is north of', for example, would
typically be formalized as a two-place predicate. We might, for instance, write
'Knoxville is north of Atlanta' as 'Nka'.
Some predicates have three or even more places. The English phrase 'is be-
tween' links three names, as in the sentence 'Nashville is between Memphis and
Knoxville', which may be written as 'Bnmk'. There is no overall convention gov-
erning the order of the names in such cases; what matters is that within a given
context we keep the order consistent.
Predicates with two or more places are called relational or polyadic predi-
cates. Those with only one place are, by contrast, said to be nonrelational or
monadic. Sentence letters are, as we shall see in the next section, sometimes re-
garded as zero-place predicates.
One final word about formalization Keep quantifiers as close as possible to
the uariables tbey gouern. This prevents mistakes. Consider, for example, the sen-
tence 'All lovers are happy', which is correctly formalized, using'IJ for'lovers'and
'H' for 'is happy', as 'Vx(3ylry - Hx)'-that is, "for all x, if there exists a y such
that x loves y, then x is happy." Here the existential quantifier is inside the brackets,
taking the narrowest possible scope-as it should be. (Since the universal quanti-
fier governs the variable 'x', which occurs in both the antecedent and the conse-
quent of the conditional, it must contain the whole conditional within its scope.)
Sometimes beginners in predicate logic are tempted to draw the existential quan-
tifier out of the conditional, like this: Yxly(Lxy - Hx).This is wrong! tvhat it
says is that for any r there exists a y such that if x loves y then r is happy. This,
given the meaning of the material conditional, is true in any possible situation in
which each x loves nothing. It does not even come close to saying that all lovers
are happy.
This mistaken formula exhibits another warning signal that should have
prevented us from writing it in the first place. It contains an existentially quantified
conditional, and existentially quantified conditionals, as we noted earlier, have
meanings so strange that they are virtually never what we intend to say.
Exercise 5.1
Using the following interpretation, formalize the arguments below in predicate
logic. (All are valid.)
One-Place Two-Place
Names Sentence Letters Predicates Predicates
a-Al H
- happiness is A-isanact L
- loves
maximized
b Beth B is blameworthy R respects
- - -
c Carl G is good
- -
(continued)
17O Csapren 6
One-Place Two-Place
Names Sentence Letters Predicates Predicates
h healing F is fortunate
s
- sleeping J
- is just
t - theft P
- is praiseworthy
- -
1. Beth is fortunate. So is Carl. Therefore both Carl and Beth are fortunate.
2. Theft is not praiseworthy, but blameworthy. Hence theft is blameworthy.
3. Healing is good. So something is good.
4. Healing is a good act. Therefore some acts are good.
5. Everything is good. Consequently theft is good.
6. Theft is not good. Therefore not everything is good.
7. Everything good is praiseworthy. Healing is good. Therefore healing is
praiseworthy.
8. All just things are praiseworthy. All praiseworthy rhings are good. Therefore
any just thing is good.
9. Healing is a praiseworthy act. Nothing praiseworthy is blameworthy. There-
fore, healing is not blameworthy.
10. Healing is good, but theft is not. Therefore some things are good and some
things aren't.
11. Everything is good and everything is just. Therefore everyrhing is both good
and just.
12. Either everything is praiseworthy or everything is blameworthy. So everything
is either praiseworthy or blameworthy.
13. If everything is good, then happiness is maximized. So, since happiness is not
maximized, not everything is good.
14. Not all acts are just. Therefore rhere are acts that are not just.
15. Not every act is just. Hence some acts are unjust.
16. Some acts are just. Nothing just is blameworthy. Therefore some acts are not
blameworthy.
17. If all acts are just, then happiness is maximized. So, if happiness is not maxi-
mized, then some acts are not just.
18. Sleeping is an act, but it is not praiseworthy and it is not blameworthy. There-
fore some acts are neither praiseworthy nor blameworthy.
19. No acts are both praiseworthy and blameworthy. Theft is a blameworthy act.
Therefore theft is not praiseworthy.
20. All acts that are not good are blameworthy. No good acts are blameworthy.
Therefore acts are blameworthy if and only if they are nor good.
21. lf Beth loves and respects Al, then Al is fortunate. But then Beth does not love
Al, since Al is not fortunate, though Beth respects him.
22. Although Carl doesn't love Beth, Al does. Therefore Beth is loved.
23. Beth loves everything. Everything loves Beth. It follows that Al loves Beth and
Beth loves Al.
24. Al loves both himself and Beth, since everything loves everyrhing.
Crassrcar Pnrorcarr Locrc: Svrurax 171
In this section we formalize the language of predicate logic, as we did for the
language of propositional logic in Section 2.3. We begin by listing the character
set for predicate logic, which consists of six distinct categories of characters:
As in propositional logic, the numerals are used for subscripts, which may
be added to predicates, names, or variables. Though seldom used in practice,
subscripts are needed to ensure that the language contains enough symbols for
very complex problems.
An atomic formula is a predicate followed by zero or more names. (A predi-
cate followed by zero names is just a sentence letter, as in propositional logic.) If
a As with the propositional operators, there is some variety in the notation for
quantifiers. Sometimes brackets are used so that'for any x' is written as '(Vx)'and
'for at least one x'as '(3x)', but the brackets are superfluous. When brackets are
used, the symbol'V'may be omitted so that'for any x'is written simply as'(x)'.
172 CHqprEn 6
'(lxFx & Gr)' are therefore not formulas. (In the last of these, the variable at-
tached to 'G' has no corresponding quantifier, since the scope of the existential
quantifier is just'lxFx'.)
Moreover, in a formula each variable occurs at least twice-once following
its quantifier and at least once again where it has replaced a name. Thus, for
example, 'lxLaa'is not a formula, since 'r' does not have a second occurrence.
The phrase 'by some variable B not already in @' in rule 4 ensures that
quantifiers of the same variable never have overlapping scopes. So, for example,
'Laa' is a formula by formation rule 1, from which it follows by rule 4 that'1xLxa'
is a formula, but we may not conclude from this by another application of rule 4
that 'lx1xLxx' is a formula, since 'rc' already occurs in '1xLxa'. Rule 4 does,
however, license the conclusion that 'lJyi.xLxy' is a formula, since 'y' does not
already occur in '1xLxa'. Quantifiers of different variables, then, may have over-
lapping scopes.
Two quantifiers of the same variable may occur in the same formula, pro-
vided their scopes do not overlap. For example, by formation rule 1, both'Fa'and
'Gb' are formulas. Applying rule 4 to each of these, we deduce that 'VxFx' and
'lxGx'are both formulas. And then, applying rule 3 to these two, we obtain the
conclusion that '(VxFx & 3xGr)' is a formula. In this formula there are two
quantifiers of 'x', but their scopes do not overlap.
The formation rules are used, as in propositional logic, to verify that formal-
izations are grammatical-that is, that they make sense in the language of predi-
cate logic. Take, for example, the expression'1xYy(Fxy * -Gr)'. We can see that
this is a formula as follows:
Exercise 6.2.1
Some of the expressions below are formulas; others are not. For those that are, tell
how they are constructed from the formation rules, as in the example immediately
above. For those that are not, explain why they are not.
l. YxLxx
2. lxYxLxx
3. laFa
4.1xFa
5. VrVY(Lr! * LYx)
6. (Laa)
7. -Yx-Fx
8. (3xFx * frFx)
9. (P -* 3rFr)
10. Lab* Lba
174 CHnprrn 6
Exercise 6.2.2
Formalize the sentences below, using the indicated interpretation, and check your
formalization with the formation rules to make sure that it results in a genuine
formula.
One-Place Two-Place
Names Predicates Predicates
Aristotle F-isafeminist
a
- R
- ridicules
n Nietzsche G is Greek
- - S
- is smarter than
'W
p Plato P-isaphilosopher
- - wrote
L. Aristotle is Greek.
2. Plato is a Greek feminist.
3. If Plato is a feminist, then someone is a Greek feminist.
4. No Greeks are feminists.
5. All feminists are philosophers.
6. All Greek feminists are philosophers.
7. Aristotle wrote something.
8. Aristotle wrote everything.
9. Aristotle wrote nothing.
10. Nietzsche ridicules everything that Plato wrore.
11. Nietzsche ridicules all feminists.
12. Nietzsche ridicules everyone smarter than he is.
1"3. Some Greeks are philosophers and some are not.
14. Some Greeks both are and are not philosophers.
15. Nietzsche ridicules a thing if and only if it does not ridicule itself.
16. Some philosophers ridicule themselves.
17. Some philosophers ridicule everything.
18. All things that are ridiculed ridicule their ridiculers.
19. If one thing is smarter than a second, then the second is not smarter than the
first.
20. Nietzsche ridicules all Greek philosophers.
6.3:DENTITY
Another logical operator which is often employed in predicate logic is'=', a symbol
familiar from mathematics. But unlike any logical operator we have so far consid-
ered, '=' is a predicate-a two-place predicate, to be precise. What makes it a
logical operator, rather than an ordinary predicate, like 'F' or 'IJ, is that its mean-
ing is fixed. It always means "is identical to" or "is the same thing as," whereas
other predicates are interpreted to mean different things in different contexts. Also,
Crassrcal PRrorcarE Locrc: Svnrax 175
Any predicate followed by zero or more names or any formula of the form
a - B, where ct and B are names, is an atomic formula.
The formation rules then remain the same. The identity predicate in combination
with the other operators enables us to express a wealth of new concepts.
Eise, Other Than, Except: The expressions 'else', 'other than', and 'except' often
express the idea of difference, that is, nonidentity. The statement 'God is more
perfect than anything else', for example, can be symbolized as 'Vx(-x = I - PBx)',
reading 'g' as "God" and 'P' as "more perfect than". Notice that the statement
'God is more perfect than anythin g' , 'Y xPgx' , is absurd, since it implies that God
is more perfect than himself, that is, 'Pgg'. The notion that God is more perfect
than anything else is inexpressible without the identity predicate.
The terms 'other than' and 'except' work similarly. The statement'AI will go
with anyone other than Beth', for example, may be symbolized as 'Vx(-x = b -
Gax)' . This formul a may also symbolize the sentence 'Al will go with anyone
except Beth'. 'Except', however, often carries an implication of exclusion so that
'Al will go with anyone except Beth' may mean not only that Al will go with
anyone who is not Beth, but also that Al will not go with Beth. If so, then a better
formalization is 'Vr(-x = b - Gar) & -Gab'-or, more compactly, 'Vx(-x = b *
Gar)'. [This compact formulation is equivalent to the first, since '-Gab'is equiva-
lent to'Vx(Gax + -x = b)'.] This might even be what is meant in some contexts
by the sentence'Al will go with anyone other than Beth'. But neither'other than'
nor 'except' always carries this implication of exclusion. If, for example, we know
that Al will go with anyone who is not Beth but are unsure of whether he would
go with Beth, we might still say, "Al will go with anyone except Beth," without
176 Cunprrn 6
Superlatives. Expressions such as 'the greatest', 'the fastest', 'the most expen-
sive', 'the most perfect', and so on are called superlatives. They are used to denote
the highest degree of a comparative quality. The comparative expressions associ-
ated with the superlatives just mentioned are, respectivelS 'greater','faster','more
expensive', and 'more perfect'. Superlatives can be analyzed in terms of compara-
tives together with the identity predicate. To say that Al is the fastest runner, for
example, is to say that Al is faster than all other runners-that Al is a runner who
is faster than all runners who are not Al. Using 'a' for 'Al', 'R' for 'is a runner',
and 'F' for 'is faster than', this goes over into predicate logic as 'Ra & Vr((Rx &
-x = a) - Fax)'. Other statemenrs conraining superlatives may be analyzed simi-
larly. The statement 'There is no largest number', for example, may be written as
'-3x(Nx & Vy((Ny & -y = x) Lxy))', reading 'is a number' for 'N' and 'is larger
than'for'L'.
-
At Least. Since the existential quantifier means 'for at least one', the formaliza-
tion of the expression 'at least one' is simple. The statement 'There is at least one
mind', for example, is formalized as 'lxMx', where 'M' means "is a mind." But
what about 'There are at least two minds'? An obvious suggestion is '3r3y(Mx &
My)'. This, however, will not do, because nothing guarantees that the variables'f
and'y' designate different objects. In facr, this is just a redundant way of saying
lxMr,' the two formulas are equivalent. To guarantee that x and y are distinct, we
need to say so. A proper formalization is '3x3y((Mx & My) & -x = y)'. Similarly
we can formalize'There are at least three minds'as'3r3y3e(((Mr & My) Sc Me)
& ((-x = y U -y = z) U -x = z))', and so on.
At Most. Suppose we want to say that there is at most one mind. This means
that no two distinct things are minds; that is, if we choose any ob;'ect r and any
Crnssrcal Pnrorcarr Looc: Svrurax 177
object y and they turn out both to be minds, then they are identical. Thus one way
to formalize this statement is 'VxVy((Mx Er My) + x = y)'. Another way to go at
it is to notice that'There is at most one mind' is the denial of 'There are at least
two minds'. This produces the formalization '-frly((Mx & Iuty) k -x = y)'. The
two formalizations are equivalent and equally correct.
But both formalizations are ungainly for formalizing'at most'with numbers
larger than one. 'There are at most two minds', for example, is 'VrVyVe((Mr &
My) 6c Mz) * ((x = y v ! = z) v x = z))' by analogy with the first method,
'-3x3y3z(((Mx & My) & Mz) & ((-x = y U -y=z) & -x = z))' by analogy with
the second. The first formula says that for any things x, y, and z that are minds, at
least two of them are identical. The second says that there do not exist three
distinct minds.
A simpler and more subtle formalization of 'There are at most two minds' is
'lx1yY z(Mz * (7 = x v z = y))' . That is, there are objects x and y such that anything
that is a mind is one of these. This does not assert that x and y are distinct (so they
might be the same thing), nor does it assert that they are minds (only that if there
are minds, then they are identical with either x or y). Thus it is compatible with
there being one mind or no minds, as well as two. But it is not compatible with
there being three, since it asserts that every mind is one or the other of x or y' It
is thus equivalent to both of our earlier formulations of 'There are at most two
minds'.
Using this more subtle formalization scheme, 'There is at most one mind' is
simply '3rVy(My - ! = x)',and'There are at most three minds' is'1xly1ZYw(Mw
- (kl = x v w = !) v w = z))'. This is the simplest way to express'at most'in
predicate logic.
We can also say there is at most one thing-where 'thing' is not a predicate
but a place-holder for a variable. (Certain mystics hold something like this.) One
simple formalization is 'lxVy ! = x' . Can you think of others?
Numerical Quantifiers. How can we say there is exactly one mind? This means
that there is at least one mind and at most one mind so that one formalization is
'lxMr & 3xVy(My - y = x)'. A more compact way of saying the same thing is
'3xVy(My * y = x)'. This combines into a single quantified biconditional the
statement 'there is at least one mind'-which is normally just '3xMx' but may be
equivalently written as '3xVy(y = x + My)'-and 'there is at most one mind'-
'3rVY(MY-Y=x)'.
Generalizing, we may formalize 'There are exactly two minds' as '3x3y(-x =
y t*.YzlMz * (z= x y 7= y)))'. This formula says that there are two distinct
objects x and y and that a thing is a mind if and only if it is identical to one of
these. 'There are exactly three minds' is '!xly3z(((-x = | & -y: z) E< -x = e) &
Yw(Mw * l(w = x v w = yl v w= e)))'-and so on. In this way we can formalize
numerical quantifiers, which are phrases of the form 'there are exactly n . . .',
where'z' stands for a number.
it true or false-and what makes it so? Bertrand Russell's consideration of this and
similar examples led him early in this century to the theory of descriptions that
bears his name and still stands out as a paradigm of logical analysis. A definite
description, according to Russell, is a phrase of the form 'the so-and-so'. Examples
are 'the present king of France', 'the positive square root of two', 'the supreme
being', and 'the woman I met last week'.
Russell's quarry was the logical form of sentences containing definite descrip-
tions. Superficially, a sentence like 'The present king of France is bald' seems to be
of subject-predicate form, with the definite description 'the present king of France'
functioning as a name. Thus we might symbolize it as 'Bk', where 'B' means "is
bald" and 'k' means "the present king of France." But difficulties ensue, for if 'the
present king of France' is a name, what does it name, and what determines whether
or not what it names is bald? These questions lead into dim metaphysics. (I7hat
sort of entity is the present king of France, anyway?)
Logical paradox is not far behind. Presumably the srarement 'The presenr
king of France is bald' is false-from which it follows that 'The present king of
France is not bald', that is, '-Bk' is true. But there is no present king of France of
whom it can be said that he is not bald. Hence 'The present king of France is not
bald' also seems false. Is '-Bk', then, both true and false?
The problem, according to Russell, lies in thinking that the logical form of
'The present king of France is bald' is 'Bk'. Definite descriptions, he argues, are
not simple names; sentences containing them have a logical structure quite differ-
ent from their surface grammar. A statement of the form 'The F is G' means
This, as we expect, is false, since the first conjunct claims falsely that there is a
present king of France. Since (A) is false, its negation,
that is,
(i) It is not the case that there is at least one present king of France, who
alone is presently king of France, and who is bald,
is true. This accounts for the intuition that'The present king of France is not bald'
is true. But what of the intuition that this same sentence is false? That is also right,
according to Russell! For the sentence is ambiguous; it can also mean
(ii) There is at least one present king of France, who alone is presently king
of France, and who is not bald,
that is, in symbols:
(C) lr((Kr & VY(KY - Y = x)) & -Bx)
Statement (ii) and its formalization (C) are both false for the same reason that (A)
is: There is no present king of France. The difference between (B) and (C) lies in
the scope of the negation operator. The ambiguity is therefore a matter of scope,
like the ambiguities that arise from the mixing of universal and existential quanti-
fiers in English. Formalization reveals the ambiguity and untangles conflicting
intuitions; the English sentence that appears to be both true and false is from a
'S7hat
logical point of view really two sentences. troubled us, then, was merely a
grammatical illusion.
In summarS Russell shows how certain sentences containing expressions
that seem to refer to nonexistent entities can be analyzed into formulas containing
only well-understood quantifiers and predicates. This both dispels logical paradox
and forestalls a metaphysical snipe hunt.
Exercise 6.3
Formalize each of the following arguments using the interpretation indicated be-
low. The arguments, by the way, are all valid.
One-Place Two-Place
Names Predicates Predicates
a一 Al C-isacarpenter B is a brother of
-
b一 Bud D-isadoctor L loves
-
c一 Cindy H is happy R is richer than
- -
d一 Dcb S-isasmoker
c一 Ed W writes novels
-
180 Cuaprun 6
6.4 FUNCT10NS
different person. If, for example, we prefix it to the name 'Martin Luther King, Jr.'
to produce the expression 'the father of Martin Luther King, Jr.', this new expres-
sion is a complex name denoting the individualMartin Luther King, Sr. 'The father
of is a one-place function symbol.
Or, to take a different example, consider the expression 'the sum of . . .
and . . .', which is represented in mathematics by the symbol '+'. Applied to the
names of two numbers, this symbol produces a complex name which denotes a
third number. Thus the expression 'the sum of two and three'-or in mathemati-
cal notation'2 + 3'-is a complex name for the number five. Thus either expres-
sion-'the sum of . . . and . . .'or'a'-is a two-place function symbol.
In both natural and formal languages functional notation is quite varied.
Some function symbols are written before the names to which they apply. An
example is the mathematical symbol '-', used as a one-place function to mean
"negative." Such function symbols are said to be written in prefix position. Other
function symbols, such as 'p' (which, when applied to a numeral, yields the name
of a price in cents), are written after the names to which they apply; these are said
to be written in postfix position. Two-place function symbols, like '+', are often
written in infix position-that is, between the names to which they apply.
The objects denoted by the names to which a function symbol is applied on
a specific occasion are called arguments of the function.5 The object denoted by
the complex symbol consisting of the function symbol and the names to which it
applies is called the value of the function. Thus for the expression '2 + 3' the
arguments are 2 and 3 (in that order) and the value is 5.
Corresponding to each function symbol is a domain of objects to which the
symbol applies and a range of objects which the resulting complex names denote.
The domain is the set of possible arguments and the range the set of possible
values. The domain of the expression 'the father of is the set of all sons and
daughters; its range is the set of all fathers. The domain of '+' is the set of all pairs
of numbers; its range is the set of all numbers.
The range and domain may contain very different sorts of objects. For ex-
ample, the expression 'the birthday of is a one-place function symbol which when
applied to the name of a person produces a complex expression that names a date.
Thus its range is the set of people and its domain is a set of dates. Here are some
examples of function symbols, along with the kinds of objects that constitute their
domains and ranges:
Number
Function Symbol of Places Domain Range
5 This is a wholly new use of the term 'argument'. It has nothing to do with the by
now familiar use, according to which an argument is a set of premises together
with a conclusion.
182 CuaprEn 6
Number
Function Symbol of Places Domain Range
e numbers prlces
Vrlrm(x)
Crassrcar PRrorcnrr Locrc: Svrurax 183
which, literally transcribed, means "For all x, r loves the mother of x." (We here
assume a domain of people.) Here are some further examples, using 'm' for 'the
mother of','l: for'loves', and'o' for'Olaf':
Sentence Formalization
Exercise 6.4
Translate the following arguments into predicate logic with functions. Assume that
the domain of discourse is the set of nonnegative integers {0, L,2,...}. (Al1 of
these arguments are valid-and sound!)
One-Place Two-Place
Function Function One-Place Two-Place
Name Symbols Symbols Predicates Predicates
o― zcro E一
f
- the square of
p the product ls evcn G is greater
-
of (i.e., "times")
-
than
l一 one C)― is odd
s
- the successor
of (i.e., the next L is less than
number after)
-
+oFtr
1. The square of zero is evIn. Therefore something is eve-n.
2. The square of one is odd. Therefore the square of something is odd.
3. Zero times the successor of zero is zero. The successor of zero is one. One is
odd. Therefore there is an odd number whose product with zero is zero.
184 CHapren 6
4. The successor of any odd number is even. One is odd. Therefore the successor
of one is even.
5. For any two numbers x and y, the product of x and y equals the product of y
and x. The product of zero and any number is zero. Therefore the product of
any number and zero is zero.
6. The square of any number is its product with itself. Zero times any number is
zero. Therefore the square of zero is zero.
7. For any two numbers, the first is greater than the second if and only if the
second is less than the first. The successor of a number is greater than that
number. Therefore every number is less than its successor.
8. The successor of zero does not equal the successor of the successor of zero.
Hence zero does not equal the successor of zero.
9. Every number is distinct from its successor. Thus there are at least two things.
10. One is the square of one. One is odd. No odd numbers are even. Therefore
not all square numbers are even.
ll. Zero is even. So every number has a successor.
12. Since any number's successor is greater than that number, there is no greatest
number. For no number is its own successor, and if one number is greater than
another, the second is not greater than the first.
C〃 ハ ρ
「
fR
CLASSICAL PREDICATE
LOGiC:SEIVIANTICS
This chapter introduces the classical semantics and the tree test for
predicate logic. As a preliminary, however, we begin with a little
elementary set theory, since set theory is the basis for the semantics.
' One of the best introductions to elementary set theory is Robert R. Stoll, Set The-
ory and Logic (San Francisco: 17. H. Freeman ,1961).
185
185 Cuaprrn 7
Set identity is also insensitive to the ways in which the members of the sets
are described or named. So, for example, the set {2, 5} is the same set as the set
{1 + 1, 5}. The fact that the number two is described in two different ways makes
no difference to the set.
It is, we should admit, a bit misleading to say that a set is a multiplicity of
objects, since in some cases a set has only one member and in one case-the empty
or null set-it has none. Sets with only one member are called unit sets or single-
tons. The set {Socrates}, for example, is the unit set of Socrates-that is, the set
whose sole member is Socrates. However, {Socrates} and Socrates are not the same.
The former is a set, the latter an Athenian philosopher. The former has one mem-
ber, Socrates; but Socrates-who is an individual, not a set-has no members in
the set-theoretic sense.
We shall use the obvious notation '{ }' to designate the empty set, though
often '0' or 'A' are used. Since sets having the same members are identical, there is
only one empty set-for any two empty sets, having the same members (i.e., no
members at all), must be identical.
The lowercase Greek letter epsilon, 't', is used in set theory to mean "is a
member of." Thus '1 e {1,,2,3}' and'Socrates e {Socrates}' are true statements, but
'0 e {1, 2,3}' and 'Socrates e { }' are false. In general, to indicate that it is not the
case that A e B, we may write 'A C B'.
Sets may contain other sets. The set I{1,2}, {3}}, for example, has two mem-
bers, the sets {1, 2l and {3}. Thus {1,2)e{11,21, {3}}and {3}e {{1, 2},{3ll,but, for
example, 1 g {{1, 2}, {3}},though 1t {1,21.
Sets which are infinite or too big to specify as lists may usually be defined by
one-place predicate phrases. For example, the set of all mathematicians is defined
by the phrase 'is a mathematician'. For this, it is common to use the notation'{xlr
is a mathematician]', which is read as "the set of all x such that r is a mathemati-
cian." Thus, for example, since Descartes is a mathematician, the statement'Des-
cartes t {xlx is a mathematician}'is true.
Another way to specify infinite sets, if their members have a natural order, is
to list the first few members within curly brackets and follow them with three dots,
indicating that the list is to be continued infinitely. Thus, for example, we may
denote the set of all natural numbers (whole numbers greater than 0) either as '{rlr
is a natural number]' or as'{1, 2,3, . ..}'. However, in the second case the ordering
serves only as a convenient way to indicate which objects we are talking about. It
does not mean that the integers are ordered within the set.
If we wish to consider ordered lists (provided these are 6nite), we shall use
instead the concept of an n-tuple. The simplest example of an z-tuple is an ordered
pair. An ordered pair is (unsurprisingly) a pair of objects taken in a certain order,
but without regard to how the objects are described. For example, the ordered
pair consisting of the numbers one and two, in that order, is different from the
ordered pair consisting of the numbers two and one, in that order. 'We denote
ordered pairs by listing their members in order and separated by commas in
between angle brackets. Thus the first of these pairs is <1, 2> and the second
is <2, 1>.
The manner of description of the objects makes no difference to the identity
of the pair. So, for example, given that x - rt,11, xr) is the same pair as <1,2>.
Crassrcal PRrorcarr Locrc: Seuarurtcs 187
{<1, 2>, 11, 3), <2, 3>, 11, 4), <2, 4>, <3, 4>, . . .}
This one is the less-than relation, since it is the set of all pairs of natural numbers
such that the first item in the pair is less than the second.
Or, to take yet another example, this finite set of ordered triples of truth
values is a three-place relation:
It is the set of all triples of truth values such that the third item of each pair is the
truth value of any conjunction whose two coniuncts have, respectively, the first
two items as values. We may therefore think of this three-place relation as the
conjunction relation.
There are also nameless random or arbitrary relations, such as this three-
place relation on the natural numbers:
Exercise 7.1
Tell which of the following statements are true and which are false, and explain
why'
L. 1L,2l = {2, 1.1
2. {1.,21={2,1,2}
3. {1}=1
4. <1,2>=<2, 1>
188 CHapren 7
5.<1,1>=1
6. If χ=1,then{1,χ }={1}.
7. If χ=1,thcn<1,χ >=<χ ,χ >.
8。 18{1)
9. <1,2>8{<1,2>,<2,4>}
10。 18{<1,2>,<2,4>}
11.{χ =1}={1}
lχ
12.18{χ iS a number}
lχ
13.{1,2}ε {χ lχ is a numbcr)
14.{<2,3>}is a twO― placc relatiOn
15。 {<1>,<2>}={1,2}
In laying out the semantics for propositional logic, we assumed that the logical
meaning of a term is given by the truth conditions for sentences containing it. We
retain that assumption here. Therefore our task in this section is to describe truth
conditions for sentences containing names, predicates, and quantifiers.
\We shall divide this task into three components: first,
an informal description
of the semantics of predicate logic; second, an informal account of two kinds of
interpretations for this semantics; and finally a detailed formalpresentation of the
semantics.
\We begin
our informal description of the semantics with a look at simple
subject-predicate sentences. Take the sentence'Bertrand [Russell] is a philosopher',
which we may symbolize as 'Pb'. The following points seem fairly obvious:
1. The name 'Bertrand' denotes an individual object: the man Bertrand
Russell.
2. The predicate 'is a philosopher' in effect denotes a set: the set of all
philosophers.
3. The sentence 'Bertrand is a philosopher'is true in a given situation if
and only if the object denoted by the name in that situation is a member
of the set denoted by the predicate in that situation.
Since Bertrand is in the actual situation a member of the set of all philosophers,
principle 3 implies (correctly) that the sentence 'Bertrand is a philosopher' is actu-
ally true. But we are talking here of the English sentence. lVhat of the correspond-
ing formula 'Pb'?
In propositional logic, the sentence letters have no truth value of their own
but are assigned various truth values on various valuations. Moreover, their inter-
pretation is not fixed but varies from problem to problem. \We specify an interpre-
tation by associating sentence letters with ordinary statements. A valuation and an
intepretation together define a possible situation.
The same thing is true in predicate logic, but valuations and interpretations
are somewhat more complex, for, like sentence letters, the names and predicates
Crnssrcar PReotcerr Loctc: Srunnrtcs 189
of predicate logic have no values on their own and (except for the identity predi-
cate) no fixed interpretation. Their values are assigned by a valuation. But these
values cannot be truth values, for predicates and names are neither true nor false.
Rather, as items I and2 above suggest, the value of a name is an object and the
value of a one-place predicate is a set. Hence a valuation for predicate logic assigns
objects to names and sets of obiects to one-place predicates.
The set assigned to a one-place predicate may be empty. For example, the
predicate phrase 'is an even prime number other than 2' applies to nothing at all
and hence denotes the empty set. Thus, if we wanted to represent this predicate
phrase by a corresponding predicate in predicate logic, we would assign to that
predicate the empty set.
As item 3 suggests, a valuation for predicate logic assigns truth conditions of
the following sort to simple subiect-predicate sentences:
The truth conditions for atomic formulas containing predicates of more than
one place (relational predicates) are only slightly more complex. Consider the
relational predicate'was a teacher of'. This predicate is true of just those pairs of
individuals such that the first was a teacher of the second. So, for example, it is
true of <Socrates, Plato>, but not of <Plato, Socrates>, since Socrates was Plato's
teacher, but not vice versa. If we suppose that Socrates was his own teacher, this
predicate is also true of the pair <Socrates, Socrates>. In general, then, a two-place
predicate denotes a set, not of individuals, but of ordered pairs of individuals-
that is, a two-place relation.
Accordingly, a valuation for predicate logic assigns to each two-place predi-
cate a two-place relation. And the truth conditions for such a sentence look some-
thing like this:
\What has just been said of two-place predicates can be generalized to all relational
predicates. To each three-place predicate a valuation assigns a set of ordered tri-
pl.r, to each four-place predicate a set of ordered quadruples, and so on-and the
iruth conditions for sentences formed from these more complex relational predi-
cates function analogously.
Like one-place predicates, relational Predicates may be empty. If we assign
the empty set to a relational predicate, that indicates that nothing in the domain
has the corresponding relation to anything. If the predicate means "loves," for
example, assigning it the empty set means contemplating a loveless situation.
190 Cuapren 7
messy', for example, is true in a situation whose domain consists just of Dolly,
Bear, Truck, and Ball iff at least one of the four statements listed above is true. The
domain delimits the set of objects among which the 'something' of this statement
is to be found.
Mixing predicate logic and English we might say that 'lx x is messy' is true
in a situation iff some statement of the form 'a is messy' is true in that situation,
where 'a' is a name of some object in the situation's domain. Likewise, 'Vr r is
messy' is true in a situation iff every statement of the form 'a is messy' is true for
any name 'a' of an object in the relevant domain.
Indeed, given this domain, we can think of 'vx r is messy' as shorthand for
the conjunction:
Dolly is messy & Bear is messy 6c Truck is messy & Ball is messy.
This analysis eliminates quantifiers altogether, reducing them to the familiar op-
erators oi propositional logic. The Austrian philosopher and logician Ludwig
'Wittgenstein advocated just such a reduction in his influential work Tractatus
Logio-Pbilosophicus (1,921). But this reduction depends on two assumptions that
are not always true:
1. Each object in the domain has a name (for otherwise we cannot formu-
late the conjunction or disjunction into which a quantified statement is
to be analyzed), and
2. The number of objects in the domain is finite (for otherwise the conjunc-
rion or disjunction into which the quantified statement is to be analyzed
has infinitely many conjuncts or disjuncts, and we have not given mean-
ings to such infinite statements).
The values assigned to predicates and names are called their extensions. The
concept of an extension is a generalization of the concept of denotation-a con-
cept that is most at home when applied to names. Each name denotes a particular
individual. But we can also think of one-place predicates as denoting sets of indi-
viduals, n-place predicates as denoting sets of n-tuples of individuals, and even
sentence letters as denoting truth values.2 This, however, stretches the ordinary
concept of denotation uncomfortably. Generally, we shall use the technical term
'extension' to designate all of these kinds of values.
The two parts of a particular valuation together comprise a structure capable
of modeling possible situations of a certain form. In fact, valuations in predicate
logic are often called models. Just as a set of toy soldiers may model a battle, a
map may model a stretch of terrain, or a configuration of electrical potentials in a
computer may model almost anything, so a valuation (that is, a domain plus
assignments of extensions from that domain to names and predicates) may model
possible situations. It usually makes little difference which objects we choose to
comprise the domain, since most often they are only stand-ins for the things they
model. \7e shall usually work with domains of numbers, but that is merely for
convenience. The important property of a domain is how many objects it contains,
for that limits the complexity of situations we can model. The domain of a ualua-
tion must contain at least one object, but tbere is no maximum. Euen infinite
domains are permitted.
A valuation becomes a model of a particular possible situation when it is
interpreted. To interpret a model, we associare formal names and predicates with
English names or predicate phrases. But we may also consider uninterpreted mod-
els. To illustrate, let's construct a valuation for the formula '(Ma & Lab) & -Fa,.
\We shall take as the domain for this valuation
the set of numbers {1,,2,3},
and we shall assign to the names'a'and'b'the numbers 1. and2, respectively.
Thus we are modeling possible situations involving three entities, two of which
have the names 'a' and 'b' and one of which is unnamed. There is no requirement
that euerytbing be named.
Further, let's assign to the one-place predicate'M'the set {1,2}. This means
that 'M' is true of 1 and 2 bur not of 3. one-place predicates may also be empty;
that is, they may apply to nothing in the domain so that their extension is the
empty set { }. Suppose we assign the empty set ro the one-place predicate 'F'; that
means that nothing in the domain (i.e., nothing in the situations we are modeling)
has the property indicated by this predicate.
Finally, let's assign the set {<1, 1>, <1,2),<2,2>} to the two-place predicate
'L . This indicates that these three pairs stand in the relation indicated by 'L
whatever that is. -
'z Gottlob Frege, who invented the modern concept of an extension while he was
inventing predicate logic, thought of truth values in this way. A true sentence, he
held, denotes The True and a false sentence denotes The False.
Crassrcar Pnrorcatr Loctc: Seuarurtcs 193
1書 塙黒 1葛 │■ ._.■■
き ■■■
■..=____._____.____→ the numbOrl
││││■ ││││││││││││││││││ ││││││││││││
denotes
■
│││││││││││
││││:││.:││││││
Al hirnseif
194 CHaprEn 7
'We
can also create models that are not merely models but the reality itself.
To do this we let the domain consist of the actual objects of our English discourse.
Thus, instead of {1, 2,3),we might select the domain {Al, Beth, Carl}. Now'a'
denotes Al himself, rather than a stand-in. This kind of interpreted model is called
a natural model.
the name'AL'
◎
7
/\
Al himself
Natural models are defined by associating English expressions with relevant pred-
icates and names and then assigning those predicates and names the objects and
sets they actually denote.
To summarize, models may be either uninterpreted or interpreted. Uninter-
preted models represent the forms of possible situations; interpreted models rep-
resent specific possible situations. There are two kinds of interpreted models:
surrogate models, in which the objects of the domain stand for something other
than themselves, and natural models, in which the objects in the domain are the
very objects denoted by the corresponding English terms.
Just as propositional logic can be done entirely by assigning truth values to
meaningless sentence letters, so predicate logic can be done entirely with uninter-
preted models. Usually, however, we will interpret our models in order to clarify
their significance.
We may now consider the formal semantics for predicate logic. Let o be any
predicate or name. Then where V is a valuation, we shall use the expression'Y(o)'
to designate the extension that V assigns to o. Using this notation, we may define
the concept of a valuation as follows:
1. A nonempty set r
of objects, which is called the domain, and
2. An assignment to each predicate or name o in that formula or set of
formulas an extension "l'(o) which meets the following conditions:
i. If o is a name, "[(o) is a member of n.
ii. If o is a zero-place predicate, Y(o) is one (but not both) of the
values T or F.
iii. If o is a one-place predicate, V(o) is a set of members of a.
iv. If o is an n-place predicate (n > 1), V(q') is a set of ordered n-
tuples of members of n.
Note once again that although the domain must contain at least one member, the
sets mentioned in items iii and iv may be empty.
Like propositional logic, predicate logic is truth-functional; that is, the truth
values of complex formulas are determined from the truth values of simpler for-
mulas by a set of valuation rules. The valuation rules for the quantifiers, however,
introduce some novelties.
Consider, for example, the formula'YxFx'.It seems initially reasonable to
suppose, as Wittgenstein did, that this formula is true if and only if Fa is true for
any name ct-that is, if 'Fa', 'Fb', 'Fc', and so on are all true, where this list includes
a statement of the form Fct for each name a. If we drop Wittgenstein's idea that a
universally quantified statement is a conjunction and allow the list to be infinite,
the resulting undersranding of the quantifier is called the substitutional interpre-
tation. But the substitutional interpretation is still problematic if the domain con-
tains unnamed objects, because it treats these obiects as if they don't exist.
Consider, for example, the valuation on the domain {1,2, 3l on which the names
'a' and'b'name l and2 respectively, 3 is unnamed, and the extension of 'F'is {1,2}.
Then Fct is true for all names ct, but 'VxFI' ought to be false, since 'F' is not true
of the unnamed object 3. The substitutional interpretation, however, would make
'VxFx'true.
To remedy this problem, we adopt slightly different truth conditions:'VrFrc'
is true if and only if
When we name an object that previously had been nameless, we expand our
valuation. So we need to define the idea of an expansion as well:
DEFINITION Let Y be any valuation, I any object in the domain of 1/, and
a any potential name of a with respect to 7. Then the expansion "l/,",,, of V
with respect to a and I is the valuation which has the same domain as V and
assigns the same extensions to the same predicates and names, but which
a/so assigns to c the extension t, ifY does not already assign o. an extension.
lf 1t abeady assigns the extension l to o, rhen 71*,.,1 is just V.
'lxFx' is true on a valuation V if and only if for some object I in the domain
of V', and some potential name ct of 4 Fct is true on 71.,.,1.
Here V1*,,r is either Y itself [if Y(cr) - t) or a valuation just like "[ except that it also
assigns the extension I to e (if V assigns no exrension to cr). This is the idea behind
rule 10 below. Rule 9 works similarly.
We may now state the valuation rules:
Crnssrcal Pne otcaru Loctc: Srl,tarurtcs 197
7(中 )(Φα
/β )=T;
7(α
α
/β
,ィ
T )(Φ )≠
loo γ (∃ βΦ )=Ti“ fOr S01me pOtential naine α Of SOme ObieCtノ in 3,
α
ノ
7(α )(Φ β )=T;
,ィ
/β )≠ T
ソ 7)(Φ α
(α
Rules 4-8 constitute the sen■ antics for propositional logic,which is thus included
in the sen.antics for predicate logic.But predicate 10gic cOntains n10re:IRules l-3
give thc senlantics fOr atornic fornlulas cOntaining nan■ es,and rules 9 and 10
prcsent the sen.antics for quantiners。
Rule l stipulates that a fornlula consisting of a one― place predicate followed
by a nan■ e is true if the extcnsion assigned to the narne(1.e.,an ObieCt)iS a rnenlber
Of the extension assigned to thc predicate(i.c.,a Set Of ObieCtS)。 (〕 OnSidCr the
fornlula`Pa',interpreted tO nlean`tAl is a predato■ ''This ineans that`Pa'is truc if
Alis a menlber ofthe set Of predatOrS一 ∝ rath軋 ifthe ObieCt n10deling Al on our
valuation is a inenlber of the Set Of ObieCtS that mOdelS the prCdatOrS・ `Pa'iS falSe,
by contrast,iff thc extcnsion of`a'is not a mcmber ofthe extension of`P'.
According to rule 2,a fornlula consisting of an π― placc predicate follo、 ved
by κnarnes is truc ifthc κ‐
tuplc consisting ofthe extensions assigned to the narncs
198 CHapren 7
(arranged in the same order as the names) is a member of the extension assigned
to the predicate (i.e., a set of ordered z-tuples). Consider the formula 'Lab', inter-
preted to mean "Al loves Beth." This means that 'Lab' is true if the pair <Al, Beth>
is a member of the set of pairs such that the first loves the second-or, rather, if
the pair of objects modeling Al and Beth on our valuation is a member of the set
of pairs that models the relation "loves". 'Lab' is false iff this is not the case.
Rule 3 says that an identity starement is rrue iff the names flanking the
identity sign have the same extension and false iff they don't. The '=' in the expres-
sion '7(ct) ='f(P)' is an expression of the metalanguage, not the identity predicate
of the object language.
In propositional logic, the valuation rules are just the rules for constructing
truth tables. But the more complex semantics of predicate logic cannot be pre-
sented in a neat tabular form. Since domains may be of any size and may consist
of any objects, each formula or set of formulas has infinitely many valuations; thus
we cannot represent all of its valuations on a finite table, as we could in proposi-
tional logic. Still, we can use the semantics to calculate the truth values of formulas
from the values (extensions) assigned to their parts, just as we did in propositional
logic.
Listed below is a set of formulas:
1. Fa&Fb
2. a=d
3. YxFx
4. )xFx
5. 3r(Fr & Gx)
6. p-.e
7'Q-VrFx
8. Lab
9. lxlyLxy
10. YxlyLxy
Now consider the following valuation 7 of these formulas:
μ l 2 l T F ロ 1
D
2 , 2 , 5 ,
3 3 >
4.
ξ”γ
O
m
a
n
〓 一
,
<
^
γ
r
>
が
一 一
Ы
γ
C
一 一
♂
γ
∈
一 一
γ
r
P
>
一 一
γ
r
Q
り
一 一
γ
∈
F
>
一 一
4 ド
γ
G
∈
︱
,
一 〓
γ
r
r
>
4︰
4︲
ぐυ
4,
Z
′4
>
,
Iilfihich of formulas 1- 1 0 is true and which is false on this valuation ? The solutions
are as follows:
1. Since 1 e 11, 2, 3],Y (' a') e 1/('F'). So by rule 1, "l/('Fa') = T. Likewise,
since 2 E {.1.,2,3), y('b') € y('F'). So again by rule 1, "f ('Fb') = T. Hence
by rule 5, 1/('Fa & Fb') = L
Crnssrcar PRrotcarr Loctc: Seulnrurtcs 199
Exercise 7.2.1
'Which
Some additional formulas are listed below. of them are true and which are
false on the valuation given on page 198? (Show your work as in the solutions to
that example.)
1. Fa*Fb
2. -a=b
3. VxGx
200 CHapren 7
4.1xGx
5. 1x-Gx
6. -lxGx
7. Yx(Fx v Gx)
8. Vr-(Fx & Gx)
9, lxLxx
10. lyYxLxy
'we
have so far ignored function symbols. Though we shall not need to use
the semantics of function symbols in this book, for completeness we consider them
briefly here.
In addition to assigning the usual extensions to predicates and names, a
valuation of a formula or set of formulas containing function symbols assigns to
each function symbol an n-place function. An n-place function 1is a set of ordered
pairs that meets two conditions:
(Notice that the one-place successor function and the two-place successor relation
discussed in Section 7.1. are in fact the same set.) Here the first member of each
pair is the argument and the second is the value for that argument.
Condition 2 of the definition of a function requires that each argument yield
a unique value. Thus, for example, since <2, 3> is a member of the successor
function, there can be no member of the form <2,r>, where x + 3.
The semantics for function symbols in predicate logic is defined by the valu-
ation rule for function symbols, which assigns an extension to a complex name by
considering the extensions of its parts. The rule is
Here Y(Q) is the z-place funcrion assigned by v to Q, and r is the value of that
function for the arguments v(o,), . . . ,T(a,). since an n-place function has only
one value for each n-tuple of arguments, there is only one object zthat the complex
name Q1ctr, . . ., d.)denotes.
Cmssrcar PRrorcerr Locrc: Sevarrtcs 201
Let's take a specific example. Suppose some valuation Y assigns to the two-
place function symbol 'f' the addition function mentioned above. Further, for the
names'a'and'b', let Y('a') = 1 and V('b') = 2. We wish to use the valuation rule
for function symbols to determine the extension of the complex name 'f(a, b)'. The
rule stipulates that
Y('f@, b)') =; iff <<Y('a'),'f('b')>, > eY('f')
But since V('a') =1,'f('b') =2, and 'f('f') = {<<1, 1>, 2),111,2),3), <<2, L>,3>,
112,2),4),111,3),4),. . .), this means that
lf('f1a,b)') =liff <<L,2>, l>t {<<1, 1>,2),111,2),3), 112, 1>,3},112,
2),4), ((1,3), 4>, . . .\
Now the only number I such that <<1,2),1> e {<<1, 1>,2),411,2>,3>, <<2,1>,
3),112,2>,4),(<1,3>, 4>,. . .] is 3. Hence
Y('f1a, b)')= 3
Or, again, suppose that 's' is a one-place function symbol to which V assigns
the successor function and that 'a' is a name such that v('a') = 2' To calculate
'll('s(a)'), we apply the rule as follows:
Y('s1a1') = tiff <<Y('a')>, l> e "l/('s')
However, since a one-tuple of a given object is just that obiect itself, it follows
that
V('s1a1') = tiff <1'('a' ), l> e V('s')
And sinceY'('a')= 2 and"l/('s') - {<1,2>,12,3),13,4), (4,5), ), we have
V('s1ay') = liff <2, l> e {<1,, 2), <2,3>, (3, 4), <4,5>, . . .}
But the only I such that <2,1> e {<1, 2>, 12,3), (3, 4>, <4, 5>, .
) is 3. Hence
V('s1a;')= l
Exercise 7.2.2
Let'a'r 'b', and'c'be names,'s'a one-place function symbol, and'f'a two-place
function symbol, and let
Y('a'1 = 1
Y(',b',1 = 2
V('c') = 3
Y('s') = {<1,2>,12,3),13,4), <4,5>,. . .}
Y('f') = {<<1, 1>,2), 111,2),3), 112, 1>r3),112,2),4),
111,3)r4>r. . .\
Use the valuation rule for functions (together in problem 5 with the valuation rule
for identity) to calculate the following values. Show your work in the way exem-
plified above.
202 CHnprun 7
1. V('s1b1')
2. V('s(s(a))')
3. Y('f1a,ay')
4. "l/('f1a, slayy')
5. V('s1b;= s')
"l/('L) = {<1,2>l
Since <1, 2> e {<1, 2>}, clearly (V('a'), y('b')> € V('U), and so by
valuation rule 2, V('Lab') = T. However, <2,1> e {<1, 2>} so that
<y('b'), "l/('a')> tY(I:). Thus by rule 2, Y('Lba') + T. Thus, since
Y('Lab T and "v' Lba ,) * T, rt follows (by definition of validity
that 'La b F Lba is not valid, tha t IS,
Tbe reductio strategy also serues for prouing tbe ualidity of formulas, as the
following example illustrates:
In this last problem, after stating the reductio hypothesis we used the valuation
rules to analyze the formula into its components until we reached a contradiction.
This is a standard strategy.
To prove the consistency of a set of formulas, by contrast' we construct a
valuation on which they are all true. The strategy here is similar to the strategy for
showing a sequent to be invalid:
This proof is easy to understand intuitively. \7e create a valuation containing two
objects, one of which is F and the other of which is not F. Clearly on this valuation,
something is F and something is not F, and so the valuation shows the set to be
consistent. We could have used Y,.',,r, instead of Y,.".,r, for our second expansion of
"[, but there is no reason not to consider two different expansions with respect to
the same name, so long as we are looking at separate formulas or subformulas
whose quantifiers do not overlap.
To prove inconsistencn it's usually best to suppose consistency and proceed
by reductio:
■ .
│I METATHEOREM:
■ .│‐ ‐‐
・
│■ .││■
‐
.
.││││
.
‐‐ ..
‐ ‐
Thc formula fVχ Fχ l&ヨ χ∼Fχ
.‐ .││.‐ .││‐
_││.‐ ││.││_│‐
. . │‐ │
lnconslstent.
││ │=‐ │ .││ │‐
'‐ is
. │■ │
■‐ │││1 11‐ ││
PR00Fi Suppose for reductio that 8ζ ヨχ∼Fχ ls consistenti that
'
■
To summarize, problems such as those we have been doing are usually most
efficiently handled by one of two general strategies:
all members of the set true. And to prove a formula contingent, use two
valuations, one which makes the formula true and one which makes it
untrue.
2. To proue sequents or formulas ualid and formulas or sets inconsistent,
proceed by reductio. Begin, as usual, by supposing the denial of the de-
sired metatheorem. Then unpack this supposition, using the definition
of its central term (e.g., 'valid', 'inconsistent'). This unpacking will yield
a statement about some valuation "l/. (For example, if the supposition is
that a certain sequent is not valid, then unpacking will yield the state-
ment that there is some valuation "l/ on which the sequent's premises are
true and its conclusion is untrue.) Now further unpack this statement
about Y, using the valuation rules to analyze the truth conditions for
complex formulas into the truth conditions of their simpler components.
The desired contradiction should eventually emerge from this analysis.
The next metatheorem follows strategy 2, but the result is a bit surprising. If
we interpret 'a' as meaning "God," for example, the metatheorem seems to assert
that it is a logical truth that God exists (more literallS that there is something
identicalto God).
from the validity of ')x x = a'we cannot legitimately infer that it is a logical truth
that God exists, for God's existence is a matter of dispute, and so 'a' cannot
legitimately be interpreted as naming God.
There are forms of logic, called free logics, which dispense with the presup-
position that all names denote-at the cost of additional complexities. Logicians
have generally found it easier to presuppose that all the names they are using
denote than to grapple with these complexities. I concur, ar least for pedagogical
purposes, which is why I don't begin with free logics. But we'll look into them in
Section 15.1.
In addition to the presupposition that each name names, predicate logic has
yet another presupposition that creates somewhat dubious valid formulas. This is
the assumption, stated in the definition of a valuation, that the domain is non-
empty. Given this assumption, there are no valuations in which nothing exists. Yet
it seems in some sense possible for nothing to exist-particularly in certain re-
stricted domains of quantification. So here again some modification may be desir-
able. That modification, too, will be discussed in Section 15.1.
Predicate logic has some other rough edges, rhe most prominent of which
involves statements of the form 'All F are G' where the extension of 'F' is the empty
set. Consider, for example, the statement 'Al[ frogs over a hundred feet tall are
green', where 'F' means "is a frog over a hundred feet tall" and 'G' means "is
green." (There are, I assume, no frogs over a hundred feet tall.) Is this statement
true or false?
The question is baffling, for English has no clear conventions to deal with
such cases. On the one hand, the fact that there are no frogs that tall seems to
make the statement false. But, on the other, since the set of all such frogs is
empty, it is true of each thing x that if x is a frog over a hundred feet tall, then r is
green-at least when we read 'if . . . then' as the material conditional-because
for each thing x it is false that it is a frog over a hundred feet tall. Thus understood,
this sentence-and, indeed, every quantified conditional whose antecedent term is
similarly empty-is true. Hence it is also true that alI frogs over a hundred feet tall
are red-and yellow, and pink, and colorless . . . ! This is the understanding that
prevails, given the conventions of predicate logic. Thus, for example, the sequent
'-lxFx l- Vr(Fx - Gr)' is valid. We leave the proof as an exercise.
It is possible to prove results of greater generality than those so far consid-
ered. The next metatheorem establishes that for any variable B, the expressions
-3B and VB- are equivaient in the sense that any formula beginning with one
keeps its truth value if we replace it with the other. (Intuitively this is right, since
both expressions mean "for no 8.") This equivalence is important; in the next
section we use it to justify the tree rule for negated existential statements.
││││││●
`│││││
Crassrcar PRrorcare Locrc: SEl,lnrrtcs 207
γ(∀ β∼Φ )
be the
Exercise 7.3
Prove the following metatheorems, using the valuation rules and appropriate
definitions:
L. 'YxFx l- Fa' is a valid sequent.
2. 'lxEx l- VrFx' is an invalid sequent.
3. '1x -x = a' is a consistent formula.
4. The formula '!r(Fx 6c -Fr)' is inconsistent
5. The formula 'YxFx v lr-Fx' is valid.
208 Cuaprrn 7
The tree test for predicate logic retains all the tree rules for propositional logic
(Section 3.3) and adds six more-two each to deal with the two quantifiers and
the identity predicate. As in propositional logic, the purpose of a tree is to display
valuations on which all the formulas of the initial list are true. However, while
trees in propositional logic display all such valuations, trees in predicate logic do
not. Formulas true on one valuation in predicate logic are true on infinitely many
others (we can always, for example, replace the objects in the domain of that
valuation by other surrogate objects), whereas in propositional logic the number
of valuations of a given set of formulas is finite.
Still, a tree in predicate logic displays all the general classes of valuations on
which its initial list of formulas might be true-and, as in propositional logic, each
open path of a finished tree can be interpreted as displaying at least one valuation
on which the initial list is true. Moreover, as in propositional logic, if all paths
close, the initial list is inconsisrent.
\We shall begin our examination of trees in predicate logic
by considering the
tree rule governing unnegated existentially quantified formulas. An existentially
quantified formula 3B@ is rrue on a valuarion Y iff there is an object z in the
domain such that for some potential name o of 4 (D"/u-the result of replacing
every occurrence of B in O by ct-is true on "[. So to display in a tree the way in
which an existentially quantified formula can be true, we give I a name a and
assert O"/p. To ensure that a is a potenrial name for 4 no matter which object z is,
we require that ct be new to the path so that no value has been assigned to it by
valuations of the initial list. Thus the rule may be srated as follows:
We shall illustrate the use of this rule by constructing a tree for the sequent '3rFr
F Fa':
Crnssrcar Pnrorcarr Loctc: Sevenrtcs 2O9
1. /1xFx Premise
2. -Fa Negation of conclusion
3. Fb 1l
Once we introduce the new name 'b' by 3 at 3, there is nothing more to be done.
The unchecked formulas are atomic, and no further rules apply. So the tree is
finished, and its one path has not closed. This indicates that the sequent is invalid.
'What
counterexample is displayed here? Recall that a counterexample is a
valuation that makes the premises true and the conclusion untrue, and a valuation
in predicate logic has two components: a domain and an assignment of extensions
to predicates and names. In general, we shall take the domain of the valuation
defined by an open path to consist just of the objects mentioned by name in the
formulas on that path. Here, for example, the only names appearing in the for-
mulas of the open path are 'a' and 'b'. Since in uninterpreted models the identity
of the objects in the domain is a matter of indifference, we'll let them be numbers.
To keep things simple and uniform, let's stipulate that in all problems the name'a'
denotes the number 1, 'b'the number 2,'c'the number 3, and so on.3 Thus, if an
open path on a tree contains the names 'a' and'b', and no others, as in our
example, the domain of the valuation defined by that path is {1, 2}.
The stipulation announced in the previous paragraph already determines the
extensions assigned to 'a' and 'b'. They are 1 and 2, respectively. To complete the
counterexample, we need only to specify the extension of the one-place predi-
cate 'F'.
In general, extensions are assigned to predicates in the following ways: For
any atomic formula that occurs unnegated on the path, if it is a zero-place predi-
cate, then the valuation assigns that predicate the extension T; if it is a one-place
predicate followed by a name, then the number denoted by that name is in the
extension of that predicate; and if it is an a-place predicate followed by z names,
then the n-tuple of numbers denoted by those names in the order of their occur-
rence is a member of the extension of the predicate. Zero-place predicates which
either do not appear on the path or appear negated on the path are assigned the
value F; and single numbers or z-tuples not explicitly included in a predicate's
extension by the occurrence of the corresponding atomic sentence on the path are
assumed not to be in that predicate's extension.
Therefore, since 'Fb' appears on the path, the number 2, which is the object
denoted by 'b', is in the extension of 'F'. And since there is no atomic formula of
the form Fct on the path for any other name ct, 2 is the only member of the
extension of 'F'. The extension of 'F', therefore, is {2}.
Accordingly, the valuation Y defined by the open path of this tree is
Domain o1y = ll,2\
Y('a') - 1
1/(,b') - 2
Y('F') = 12\
3 We could also give naming conventions for subscripted names, but we won't
bother since in practice we won't use them,
21O Cunprun 7
It is easy to see, using the valuation rules of predicate logic, that this is a valuation
on which 'lxFx'is true and 'Fa'is false and hence a counterexample.
This is, of course, an uninterpreted valuation. To give it some intuitive con-
tent, we might convert it into a surrogate valuation by taking 'a' to mean "A1", 'b'
to mean "Bethr" and 'F' to mean, say, "is female." Thus the number one is a
surrogate for Al and two is a surrogate for Beth. Then the specific possible situa-
tion it represents is a situation involving only Al and Beth in which Beth is female,
but Al isn't. And that situation is a clear counterexample to the argument
Something is female.
Al is female.
Thus we can see intuitively, as well as formallS why the sequent 'fxFx F Fa' is
invalid.
Let's now turn to the rules for negated quantifiers:
l. /-1xFx Premise
2. -Yx-Fx Negation of conclusion
3. Yx-Fx 1 -l
4. x 2,3-
Here application of -l to the formula at line 1 produces an immediate contradic-
tion and closes the path. The sequent is valid. There is no counterexample.
Notice that the annotation at line 4 specifically mentions that the - rule was
used to close the path. In predicate logic there are two path-closing rules: the
familiar rule - and a new rule, -=, which will be introduced shortly. It is useful,
therefore, to begin to get into the habit of specifying which rule we are using when
closing a path.
Cmssrcar PRrotcnrt Loctc: Srvarurtcs 211
The tree for the sequent 'Fa & Ga l- Vx(Fx & Gx)' illustrates the use of the
-V rule:
1
/Fa&Ga Premise
2
/∼ ∀χ(Fχ 8c Gχ )
Negation of conclusion
3
Fa 1&
4
Ga 1&
5
/ヨ χ∼(Fχ 8こ Gχ )
)_Y
6
/∼ (Fb&Gb) 5l
7. ∼Fb 6∼ 8こ ∼(]b6∼ 8ζ
A、 t this point ali nonatornic fornlulas are checked,and there is nothing lcft to do,
so the trec is cornpletc.'I・ his trcc has t、 vo open paths,but both rcpresent the san■ e
valuation(1/・ :
DOmain Of γ ={1,2}
γ(`a') =1
7(`b') =2
γ(`F') ={1)
γ(`G') ={1}
And this valuation is a counterexample to the sequent. If we interpret 'a' to mean
*Alr" 'b' to mean "Bethr" 'F' to mean "is foolishr" and 'G' to mean "is greedyr"
then the sequent says that since Al is foolish and greedy, everyone is foolish and
greedy. The valuation defined by the tree represents a conterexample to this infer-
ince. The counterexample is a situation involving only Al and Beth in which Al is
both foolish and greedy but Beth is neither.
The final quantifier rule is the rule for the unnegated universal quantifier.
'!7hat
makes a universally quantified statement vBiD true on a valuation V is that
each instance of the form (D"/u is true, where @"/u is the result of replacing all
occurrences of B in O by a potential name o of an obiect in the domain. This
suggests that the tree rule for the universal quantifier should enable us to produce
an instance of the quantified formula for each object in the domain.
However, in constructing trees, we construct the domain associated with
each open path as we go. Since existential formulas may introduce new names
onto the path, we cannot in all cases be sure what the ultimate domain will be
until the path is complete. (Of course, if the path closes, there is no domain to
worry about.) Therefore, even if we instantiate a universal formula for each name
currently on the path (that is, apply the V rule for each such name), so long as the
path is not finished, that is no guarantee the path contains each instance of the
universal formula required to make that formula true. New obiects may be added
to the domain later. Thus we cannot be sure that we are done using a universal
formula until the path is finished. This means that, unlike other formulas, uniuersal
formulas sbould not be cbecked when we apply their tree rule.
Ultimately, unless the path closes, we should obtain an instance of each
universal formula for each name on the path. But what if we need to apply the
212 CHnprrn 7
universal rule to a formula on a path that does not yet contain any names at all?
For an answer, we must recall the stipulation contained within the definition of a
valuation that the domain must be nonempty. There must, in other words, be at
least one object in the domain. lwe therefore introduce a new name for that object
and instantiate the quantified formula using that name.
The universal rule may thus be stated as follows:
Let's use the tree for the sequent 'Vx(Fx * Gx), VxFr F Ga' to illustrate this rule:
1
Vx(Fx Gx)
- Premise
2
YxFx Premise
3
,/Fa- Ga 1V
5
Fa 2V
6. -Fa4- Ga4-
7. x 5,6 x3,6
Since the only name on the path is 'a', this is the name used to instantiate the
quantified formulas on lines 1 and 2 by V. The results appear at lines 4 and 5.
Note that neither universal formula is checked. Since both paths close, the sequent
is valid.
The sequent 'YxFx F lrFr' provides an example of the use of V on a path
initially containing no names:
1. ∀χFχ Premise
2. /∼ ∃χFχ Negation of conclusion
3. ∀χ∼Fχ 1=
L-)
4. Fa 1V
5。 ∼Fa 3V
6. × 4r5
At line 4, since no name yet appears on the path, we choose the new name 'a' and
instantiate'YxFx'with'a'to obtain'Fa'. But at line 5, since'a'now appears af
line 4, we use 'a' again to instantiate 'Vr-Fx'. This yields a contradiction and
closes the path, showing the sequent to be valid.
Consider now the tree for the sequent'frFx & lrGr F 3x(Fx & Gr)':
Crassrcar PRrorcarr Looc: Srvarurrcs 213
/∼ ヨ
χ(Fχ 8こ Gχ ) Negation of conclusion
3
/∃ χFχ 1&
4
/∃ χGχ 1&
5
Fa 3l
6
Gb 41
7
/∼ (Fa&Ga) 7V
9
/∼ (Fb&Gb) 7V
∼Fb 9∼ 8( ∼Gb9∼ 8(
一 う0
× 6,12∼
Domain of γ ={1,2)
γ(`a') =1
7(`b') =2
γ(`F') ={1}
γ(`G') ={2}
when nlultiple quantiners are prcscnt,there lllust bc lllultiplc applications
of the quantiner rules,as thc tree for the sequcnt`∀ χ∀γ(Fχ γ―→∼Fγ χ)卜 ∼ヨχFχ χ '
11lustrates:
1
YxYy(Fxy
- -Fyx) Premise
2
/lxFxx 2--
4
Faa 3l
5
Yy(Fay
- -Fya) 1V
6
- -Faa 5V
Trees for predicate logic may also contain sentence letters, as in propositional
logic. Here is the tree for '-lxFr - P, -Fa F P':
l. /-lxFx-P Premise
2. -Fa Premise
3. -P Negation of conclusion
4. ,/--lxFx 1- P1 *
5. /1xFx 4 -- x3,4
6. Fb sl
Note the introduction of the new name 'b' at step 6. 3 requires a new name and
does not allow us to conclude 'Fa'. Thus the left path remains open. It represents
the following valuation V:
"r('b') - 2
"l/('F') =
"l/(.P,) _ F{21
The tree tests for consistency, contingency, the validity of formulas, and other
semantic properties are the same as in propositional logic. The following tree, for
example, tests the formula '-(Vr-Fx & 3rFx)'for validity:
l. YxlyCxy Premise
2. -Caa Negation of conclusion
3. /1yCay 1V
4. Cab 3l
5. llyCby 1V
6. Cbc 5l
Cressrcal PRrorcnrr Locrc: SrveNlcs 215
7 ′ nじ
/
﹁J
C
νン
νン
1∀
V
Ccd 7∃
This tree goes on forever. Each time we instantiate the universal formula at line 1,
we get a new existential formula. Then we must apply 3 to this existential formula.
But that produces a new name, and we must go back and instantiate 1 again with
that new name. This produces yet another existential formula, and the cycle begins
anew. Yet the infinite path specifies a counterexample. The counterexample is the
valuation V such that
に
つ“
う0
4,
Domain ofγ =
1
γ(`a') =
2
7(`b') =
or unchecked) on the path which results from replacing one or more occur-
rences of either of these names by the other in (D. Do not check either o = I
or (D.
The stipulation that the formula obtained should not have occurred previously,
either checked or unchecked, on the path, prevents repetition-and hence useless
work.
Notice that, Iike a universally quantified statement, an identity statement is
never checked. This is because, like a universally quantified statement, an identity
statement may be reused repeatedly as new formulas to which it is applicable are
added to the path. This tree for the sequent 'a = b, Lab F Lba' illustrates its use:
L. a=b Premise
2. Lab Premise
3. -Lba Negation of conclusion
4. Lbb 1,2=
5. -Lbb 1,3=
6. x 4,5
At step 3 we replace the 'a' in 'Lab' by 'b' to obtain 'Lbb'. Similarl6 at step 4 we
replace the 'a' in '-Lba' by 'b' to obtain '-Lbb'.
Getting the tree to close this quickly requires some foresight; many of the
substitutions of 'a' for 'b' or 'b' for 'a' that we could have made would not help in
closing the tree. However, by making all possible substitutions of this sort in all
the unchecked formulas of the tree, we would eventually hit upon one that would
close the tree. But such "blind" substitution is inefficient; intelligent planning is
better.
The negated identity rule is encouragingly simple:
Negated Identity (-=) Close any open path on which a formula of the form
∼α =α occurs
Clearly any formula of the form -o = q is inconsistent, which is what justifies this
rule. The tree for the sequent 'a = b F b = a' illustrates the rule:
L. a=b Premise
2. -b = a Negation of conclusion
3. -b=b 1,2=
4. x 3--
The sequent is valid. At step 3 we replace the occurrence of 'a' in'-b = a'by'b'to
obtain '-b = b', which closes the tree by -=.
Consider now the tree for the sequent'Fa, Fb F a = b':
Crassrcar PReorcarr Locrc: Srvnnrrcs 217
1. Fa Premise
2. Fb Prcnllsc
3. ∼a=b Negation of conclusion
NOthing more can be〈 lone hcreo Neither the identity rule nor the negatcd identity
rule applieS tO fOrmulaS Ofthe fOrm∼ α=β ,Where α and βare difた rent nameS・ SO
thc trec is inished and the sequentis invalid.Thc opcn path reprcscnts thc fO1low―
ing valuation:
DOmain={1,2)
γ(`a') =1
γ(`b') =2
γ(`F') ={1,2}
Finally,let's consider the sequent`∼ a=b,∼ b=c卜 ∼a=c':
1. ∼a=b Premise
2。 ∼b=c Premise
3。 /∼ ∼a=c Negation of conclusion
4.a=c 3--
5。∼c=b 1,4 =
6。∼b=a 2,4 =
The only identity statement here is at line 4. \We apply this statement to each
formula to which it can be applied, but still the tree does not close. Hence the
sequent is invalid.
'When
open paths contain unnegated identity statements, we must modify
our conventions regarding how open paths define valuations. The formula 'a = c'
can be true only if 'a' and 'c' name the same object. To ensure that, we adopt the
following stipulation: 'When formulas of tbe form a = g appear on an open path,
the extension assigned to each name is the lowest number assigned to any name
with uthich it is identified, either directly or uia a series of intermediate identity
statements, Names not occurring in identity statements, however, have their usual
denotations. For example, in the tree above, 'a' would normally denote 1 and 'c'
would normally denote 3. But since 'a = c' appears on the open path, we change
the denotation of 'c' to the lesser denotation 1. Thus the valuation specified by this
tree is
f げ1 2 1
l
^
〓 一
E)ornain
一
γ(`a')
一
一
γ(`b')
〓
γ(`C')
f. is checked, or
2. is an atomic formula not of the form cr = B, or
3. is the negation of an atomic formula (including formulas of the form
_61
= B), or
4. is a universal formula to which the rule V has been applied for each
name on the path, or
5. is a formula of the form 61= B which has been applied to each formula (D
(not including o = 0 itself) that contains either the name ct or the name p
to produce every possible consequence obtainable by the identity rule.
'Whenever
you produce a tree that contains open paths, check each path with this
five-item list to make sure that the tree is really finished.
Table 7,1, summarizes the six new tree rules for predicate logic with identity,
along with the valuation rules to which they correspond (for a similar summary of
the tree rules for propositional logic, see Table 3.4 in Section 3.3):
Order of Application
In propositional logic the order in which rules were applied made no difference,
except as a matter of keeping things simple. We get the correct answer if we apply
the rules correctly, no matter what the order of application. In predicate logic this
is no longer truet Order of application matters.If we apply rules in the wrong
order, we may get an infinite tree even if the tree can be closed. Consider, for
example, the inconsistent set lYxlyFyx,I.xYy-Fyx|. (To see informally that this
set is inconsistent, think of 'F' as meaning "is the father of." Then the first of the
two formulas says that everything has a father and the second that something is
fatherless.) Now suppose we test the set for inconsistency by applying the rules in
the followin g incorrect ordert
1
YxlyFyx Given
2
/lxYy-Fyx Given
3
/lyFya 1V Wrong!
4
Fba 33 Wrong!
5
/3yFyb 1V Wrong!
6
Fcb 5l Wrong!
7
/1yFyc 1V Wrong!
8
'Wrong!
Fcd 71
'We
f
can continue applying V to 1 and then to the resulting formula forever. Yet
since the set is inconsistent the tree ought to close. In fact, it does close if we apply
the rules in a different order:
1. YxlyFyx Given
2. /1xYy-Fyx Given
3. Yy-Fya 21
Crassrcar PRrorcare Locrc: Srvaurrcs 219
TABLE 7.1
The(Eorrespondence BetVVeen ValuatiOn Rules and¬ ree Rules
γ(∀ βΦ)=T iff fOr all Universal Quantification (V) If a formula of the
pOtential naines α form VB<D appears on an open path and ct is a name
Of al1 0biCCtS/in 3, that occurs in a formula on that path, write @"/u (the
γ(中 )(Φ α
/β )=T result of replacing all occurrences of B in @ by ct) at
the bottom of the path. If no formula containing a
name appears on the path, then choose some name
ct and write @'/u at the bottom of the path. In either
case, do zot check VB@.
V(Ct = P1= Fiff l/(ct)* Negated Identity (-=) Close any open path on
V(p); more to the point which a formula of the form -ct = cr. occurs.
than this general rule is this
more specific consequence
of it: "l/(a = a) = F iff V(a) *
1/(ct).
22O Cuprrn 7
4. l)yFya 1V
5. Fba 41
5. -Fba 3V
7. x 5,6
The problem with the first tree was that we never used line 2; we simply ignored
it. If, instead, we switch back and forth among the usable lines, rather than getting
stuck in some subset of them, we can guarantee that trees whose initial lists are
inconsistent will always close. (This will not, however, always prevent infinite trees
for consistenr initial lists.) The simplest way to do this is to impose a prescribed
order on the application of rules. One such order is specified below. Apply every
rule you can under item A before going to item B; apply every rule you can under
item B before going to item C; and so on. When you finish with item E, go back
and start at the beginning of the list again, and repeat this procedure until the tree
is finished.
Exercise 7.4.1
Construct a tree for each of the following sequents to decide whether it is valid. If
it is invalid, specify one of the valuations defined by an open path of its tree that
serves as a counterexample.
1. VrFr F Fa
2. l,xEx F VxFx
3. Fa l- lxFx
4. FalYxFx
5. -fr-Fr F VxFx
6. YxFxl -3x-Fx
7. Yx-Fx l- -VrFx
8. -VrFr F Vx-Fx
9. -lxlFx & Gr) F -Fa
10. Fa-Gb,Vr-FrF-Gb
11. Vr(Fr -- Gr), -fxGx l- -Fa
12. Yx(Fx - Gx), VrGr F Fa
13. VrFr - VrGr, -lxGx I lx-Fx
14. lx(Fx & Gx) llxFx & fxGx
15. VrFx v VxGx F Vr(Fx v Gx)
16. Vr(Fr v Gr) F VrFx v VxGr
L7. 1x(Fx & -Fx) F P
18. lxFx &fx-Fx F P
79. Y xY y(Lxy - Lyx), lxLax I lxLxa
20. lxlyLxyISxLxx
21. lxYyLxylYxlyLyx
22. Yx(Fx r VyGy), Fa F VrGr
23. Yx(1yGy-Fx), Ga l- VxFx
24,Fa,Gbl-a=b
25, Fa,-FbF-a=b
26. Yx(Fx * Gr), -Fal )x-x =a
27. Lab F lxly-x =y
28. lYx x=x
29. IYxYY(x=Y-Y=x)
30. F VTVYYz((x = Y & Y = z)' Y= 71
Exercise 7.4.2
\7hat follows is one of many versions of the famous ontological argument for the
existence of God:
God is (by definition) the most perfect being. Anything that exists is more
perfect than anything that does not exist. Therefore, God exists.
Formalize this argument using the name 'g' for 'God' and the two-place predicate
'P' for 'is more perfect' (hint: review the treatment of superlatives in Section 6.3).
Then use a tree to test the argument for validity. Explain what the tree test shows.
222 CHaprrn 7
1. For tbe purposes of the uniuersal quantifier rule (Y), each complex name
formed by a function symbol is a name; a patb is therefore not complete
until the quantifier rule has been applied for eacb of these names tbat
occurs on that path.In the formula 'Lom1m1o1;', for example, there are
three names:
nl(In(0)) In(0) 0
l. /1xLxf1x1 Premise
2. -Loftot Negation of conclusion
3. Laftal 1 l
N7e apply the existential rule as usual to line 1, introducing a new name ,a,. This
leaves only atomic formulas. The tree shows that in a domain containing four
objects (corresponding to the four names, 'o', 'f1o;', 'a', and .f1ay,)
there is an
Cussrcal PRrotcate Loctc: Sruarurtcs 223
6。 b=f(a) 5∼ ←→ ∼b=f(a) 5∼ ←
→
7. ∼b=f(a) 5∼ ←→ b=f(a)5∼ →←
8。 × 6,7 ∼ × 6,7 ∼
Exercise 7.4.3
Demonstrate the validity of the sequents formalized in Exercise 6.4 by construct-
ing a tree for each.
C″ ス Pア fR
CLASSICAL PREDICAITE
LOGIC:INFERENCE
Predicate logic adds three new operators to those of propositional logic:
the two quantifiers and the identity predicate. This chapter presents a
natural deduction system for predicate logic by adding six new inference
rules-an introduction rule and an eilmination rule for each of the three
new operators. We then consider the role of function symbols in proofs.
224
Crassrcal Pnrotcarr Loctc: lrureneuct 225
,Fa l- frFr,. Similarly, given that Alice likes herself, we may draw any
symbols:
of three conclusionsr that Alice likes something (namely herself), that something
(Alice again!) likes Alice, and that something (guess who) likes itself. Symbolically:
Laal lxLax
LaallxLxa
Laal lxLxx
All of these inferences, and others of a similar nature' are licensed by the 3I rule,
which may be expressed formally as follows:
Each of the three sequents above may be proved by a single step of 3I. \tr(ith respect
to the first of these sequents, @ is 'Laa', c is 'a', B is 'r', and @o/" is 'Lar'. For the
other two, cr, p, and O are the same, but @o/" is 'Lxa' or'Lrx', respectively. The
proof of the first sequent looks like this:
l, Laa A
2. lxLax 13I
The proofs of the other two sequents are the same, except for the identity of the
formula on line 2. Thus 3I may sometimes be applied to the same formula in
several different ways.
3I may also be applied to complex formulas, as in this proof of the sequent
'Fa, Ga F 3x(Fx & Gr)':
l.Fa A
Z.Ga A
3. Fa 6c Ga 1.,2 &l
4. lxlEx & Gr) 3lI
Notice that in performing lI we must add the outer brackets that had been
dropped by convention from the conjunction. If we do not do so, the result, '3rF.r
& Gx', is not a formula.
JI never introduces more than one quantifier at a time. To prove a conclusion
with two existential quantifiers, we need two steps of lI. Consider, for example,
this argument:
In symbols: 'Dab & Dba I- lx1y(Dxy & Dyr)'. Because the conclusion contains
two existential quantifiers, the proof uses two steps of 3I:
226 Cupren 8
メヽ
a a ヽ
A
χ
2
1&E
3
口 χ
l&E
・ いr
4
2ヨ I
5
∃χGχ 3ヨ I
6
However, the premise 'Fa & Ga' must be dissected into its conjuncts .Fa, and ,Ga,
before fI can be applied in order to obtain the conjuncts of the conclusion.
Sometimes the use of lI is less obvious and direct, as in the reductio strategy
used in this proof of '-1xFx F -Fa':
1. ∼∃χFχ A
2. Fa H(br∼ I)
3. lxFx 2∃ I
4. │ fxFx & -1xFx l,3 8ζ I
5。 ∼Fa 2-4∼ I
Here, since the conclusion '-Fa' is negative, we hypothes ize'Fa' for indirect proof
at line 2. Then at line 3 we obtain by 3I a statement that contradicts assumption
1. The resulting contradiction enables us to deduce the desired conclusion by
-l at
line 5.
Sometimes fI is combined in complex ways with other rules. In this proof of
'Fa v Gb F fxFr v 3xGx', for example, it is used with vE:
Crassrcal PReotclrr Loctc: lrurtRrrucr 227
1. Fa∨ Gb A
2。 I Fa H(fOr→ I)
3. I ∃χFχ 2∃ I
4。 │ ヨχFχ ∨ヨχGχ 3∨ I
5。 Fa― → (∃ χFχ ∨∃χGχ ) 2-4-→ I
6。 I Gb H(fOr→ I)
7。 │ ヨχGχ 6∃ I
8。 │ ヨχFχ ∨∃χGχ 7∨ I
9. Gb→ (∃ χFχ ∨ヨχGχ )
6-8-→ I
10. ヨχFχ ∨ヨχGχ l,5,9∨ E
The replacement here of both 'a' and 'b' by the same variable 'x' is legitimate
becausi of the nonoverlapping scopes of the two existential quantifiers.
But a quantified variable may replace only one name at a time (though it
may replace several occurrences of that name). The inference illustrated below
violates the fI rule by attempting to replace two different names:
1. Tab A
2. SxTxx 1lI (Wrong!)
From the premise that Alice is taller than Bob, it certainly does not follow that
something is taller than itself!
Exercise 8.1
Prove the following sequents, using 3I and the inference rules of propositional
logic:
1. Lab llxLxb
2. Labl3xLax
3. LaallxLxx
4, Labl)xlyLxy
5. Fa v Ga F 3r(Fx v Gr)
6. Fa v Ga F fxFr v lrGx
7. -3x1yRxy F -Rab
8. -3rlyRxy F -Raa
9. lFa-lxFx
10. -lxFr l1x-Fx
rules because, like *I and -1, it employs a hypothetical derivation. As usual, the
idea is best explained by example. Consider the argument
We may represent this as the sequent'3x(Fr & Gx) F 3x(Gx & Fx)'. This sequent
is valid. But how might we break it down into simple inferences that also have
application elsewhere? The premise asserts that some fish (i.e., at least one) are
guppies, but it does not identify this fish or (if there is more than one) these fish.
Yet their identities are not really germane to the inference; the conclusion ought to
follow no matter who or what they are. So we might as well choose some individ-
ual arbitrarily and suppose that it is one of these fish thar are guppies. r7e might
then reason as follows:
Take Al here and suppose that he is one of these fish that are guppies. Then
Al is a fish. Moreover, Al is a guppy as well. Therefore Al is both a guppy
and a fish. So some guppies are fish.
But now since we have shown that guppies are fish from the supposition (hypoth-
esis) that Al is a fish that is a guppS and since (apart from this supposition) we
have assumed nothing specifically about Al, then no matter which object or ob-
jects are actually the fish that are guppies, it must be the case that some guppies
are fish.
The conclusion '3x(Gx & Fx)' is derived twice-once hypothetically, from the
supposition 'Fa & Ga', and a second time nonhypothetically (this time by the rule
fE). fm is always the case in applications of lE. \What justifies the step of lE (the
conclusion's change of status from hypothetical to nonhypothetical) is that in the
hypothetical derivation we could substitute dny n me for the name of the represen-
tative individual (in this case 'a') and still reach that conclusion. Thus, in this
proof, no matter which individuals are both F and G (provided only that at least
one is, which we assumed at line 1), the conclusion'3x(Fx 6C Gx)' must be true.
It is crucial for the application of lE that the name of the representative
individual not appear in any assumptions or hypotheses other than the supposition
of the representative instance in which it is introduced and that it not appear in
the conclusion of the hypothetical argument from that supposition, for otherwise
there would be no guarantee that our hypothetical reasoning applies indifferently
to any of the unknown objects the representative individual represents. There
would be no guarantee, in other words, that the representative individual is truly
representative,
In annotating an inference by !E (see line 7 above), we always cite the line
of the existential premise to which it was applied (in this case line 1), together with
the lines of the hypothetical derivation in which the desired conclusion is derived
from the supposition of a representative instance of that existential premise (here,
lines 2-6).
The lE rule is stated formally below. If you have followed the discussion so
far, you should see the need for the various qualifications that make it rather ugly:
In the proof given above, 3B<D is 'lx(Fx & Gx)', B is 'r', Y is '3x(Gx & Fx)', ct is
'a', and <D"/u is 'Fa & Ga'.
In summar5 to use an existential premise ]B(p to prove a conclusion Y,
we hypothesize a representative instance @"/u of that premise and derive Y
230 CHnprrR 8
hypothetically. So long as
cr does nor appear in any other hypothesis whose deri-
vation has not ended, nor in any assumption, nor in <D or the conclusion y, we
may then end the hypothetical derivation and infer Y by 3E. As justification for
this step, we cite the line on which the existential premise 3B<D occurs and the lines
of the hypothetical derivation in which Y is derived from <D"/u.
This proof of the sequent'lxLxx I lxlyLxy' provides another example of
the use of lE:
1
ヨχLχ χ A
2
Laa H(br∃ E)
3
SyLay 2ヨ I
4
lxlyLxy 3∃ I
5
lxlyLxy l,2-4ヨ E
\We hypothesize a representative instance of the premise'lxLxx'at line 2. From
this we derive the conclusion 'f,xlyLxy' at line 4. lE then enables us to end the
hypothetical derivation and assert 'f,xlyLxy' on the strength of the assumption
'1xLxx'alone.
In the examples we have examined so far, the conclusion y obtained by 3E
is an existential formula. But in general Y may have any form. In the case of the
sequent 'lx(Fx & -Fx) F P', for example, it is the sentence letter 'P'. The premise
of this sequent, 'lx(Fx & -Fx)', is self-contradictorn and so of course it implies
anything.
1. ∃χ(Fχ で
文;∼ Fχ ) A
2. I Fa&こ ∼Fa H(brヨ E)
3。 I Fa 28こ E
4。 │ ∼Fa 28こ E
5。 I P 3,4 EFQ
6. P l,2-5ヨ E
I7e take 'Fa & -Fa' as rhe representative instance of '3r(Fr & -Fr)'at line 2.The
hypothetical derivation uses the derived rule EFQ (see Section 4.4). Notice that all
the restrictions on lE are met: 'a' does not occur in any hypothesis other than 2,
nor in any assumption, nor in 'P'.
Violating these restrictions breeds trouble. Consider, for example, this erro-
neous reasoning:
1■ つム ^
∃χ(Fχ &Gχ ) A
I Fa 8ζ Ga H(br∃ E)
j
l Fa 2 8cE
.
4
. Fa l,2-3∃ E (WrOng!)
Everything is fine down to the last step. But the conclusion y-that is, 'Fa'-of
the hypothetical derivation contains 'a', the name of the representative individual.
So this conclusion is not, as it ought to be, provable regardless of which individual
is both F and G. Hence the step of lE at line 4 is illegitimate. It is no surprise, then,
that the resulting sequenr, '3x(Fx & Gx) l- Fa', is invalid. From the assumption
Cussrcar Pneorcare Loctc: lnrrnrrucr 231
that some fish are guppies it does not follow that Al is a guppy (Al might not be a
fish at all)!
'We
also brew trouble if we use as the name of the representative individual a
name that already occurs in an assumption or in a hypothesis whose hypothetical
derivation has not ended, as in this reasoning:
1. ヨχLχ a A
2. I Laa H(for∃ E)
3。 │ ∃χLχ χ 2∃ I
4. ∃χLχ χ l,2-3∃ E (Wrong!)
Here the name 'a' occurs in the assumption. Therefore, in taking 'a' as our repre-
sentative individual at line 2, we have chosen an individual that is not truly repre-
sentative. Suppose'I-l means "loves" and'a'means "A1." Then the assumption at
line L says that something loves Al. But we are not told what this "something" is.
By supposing ar line 2 that it is Al himself (instead of a truly representative individ-
ual about whom we previously had assumed nothing), we illegitimately introduce
information about this individual's identity. This enables us to derive an exces-
sively strong conclusion at line 4. For from the premise that something loves Al, it
does not follow that Al loves himself.
If a premise begins with two existential quantifiers' then two uses of fE are
'We shall illustrate this double usage with a proof of the sequent
required.
'3x1yLxy l lylxLyx'. The premise and the conclusion of this sequent are equiva-
lent. They say exactly the same thing-which is why the sequent is valid.
1
SxlyLxy A
2
lyLay H(for∃ E)
3
I Lab H(fOr∃ E)
4
| :,1,, 3ヨ I
5
| 1yl*ty* 4ヨ I
6
3ylxLyx 2,3-5∃ E
7
lylxLyx l,2-6ヨ E
1 The hypothetical derivation from line 3 has ended by line 7 (indeed, it ends at line
5) so that the factthat its hypothesis contains'a'is of no concern when we apply
thefErule atlineT.
232 Cunprrn 8
1。 ∼∃χ(Fχ ∨Gχ ) A
2. ∃χFχ H(br∼ I)
3. Fa H(fOrヨ E)
4. Fa∨ Ga 3∨ I
5。 ヨχ(Fχ ∨Gχ ) 4ヨ I
6。 ∃χ(Fχ ∨Gχ )&∼ ヨχ(Fχ ∨Gχ )1,58d
7. ∃χ(Fχ ∨Gχ )8こ ∼∃χ(Fχ ∨Gχ ) 2,3-6∃ E
8。 -)xFx 2-7∼ I
The conclusion is negatiye, so we employ indirect proof, hypothesizing'3xFr'for
-I at line 2. Then at line 3 we hypothesize a represenrarive instance of 'frFr' for
lE. The contradiction obtained at line 6 from this second hypothesis is shown by
3E at line 7 to follow just from 'lxFx'rogether with the asssumption. This allows
us to reject'i.xFx'and affirm '-lxFx'by -I at line 8.
We conclude our discussion of lE with another proof that uses fE within an
indirect proof. Here, however, the contradiction most directly obtainable from the
representative instance contains the name of the representative individual. ltrfe
therefore use EFQ to derive a contradiction not containing this name before apply-
ing 3E. The sequent in this case is the theorem'F -Ix(Fx * -Fx)':
1
Fa * -Fa H(br∃ E)
3
lF" H(br∼ I)
4
| -r'a 2,3-■ 4P
5
I r"&-Fa 3,4&;I
6
-Fa 3-5∼ I
7
Fa 2,6-■ 4P
8
P8(│∼ P l,2-8∃ E
0
∼∃χ(Fχ ←→∼Fχ ) l… 9∼ I
Exercise 8.2
Prove the following sequents:
1.∃ χ(Fχ 8c Gχ )卜 ∃χFχ &:ヨ χGχ
2.∃ χFχ 卜∃χ(Fχ ∨Gχ )
3.∃ χFχ ¨
→く3a卜 Fb‐ → ヨχGχ
∼∼
4. ヨχ Fχ トヨχFχ
Cllssrcal Pnrorcarr Looc: lrureREnce 233
5。 ∃χFχ 卜∃χ∼∼Fχ
6. ∼ヨχFχ 卜∼ヨχ(Fχ (文 Gχ )
7. 卜∼ヨχ(Fχ 5∼ Fχ )
8. ‐ `致∃ッFγ
トヨχFχ ・
9. ∃χ(Fχ ∨Gχ )卜 ∃χFχ ∨∃χGχ
10。 ヨχFχ ∨∃χGχ 卜∃χ(Fχ ∨Gχ )
In the inference above, VB(D is 'Vr(Sr & Tr)', (D"/u is 'Sa & Ta', B is 'x', and
c is 'a',
The application of VE is in most instances simple and obvious. Here it is
used in a proof of the sequent 'Vr(Fr * Gx), Fa l- Ga':
1. Vr(Fx - Gx) A
2.Fa A
3. Fa-Ga 1VE
4. Ga 2,3-E
As with the other quantifier rules, more than one use of VE is required to
eliminate more than one quantifier. This proof of the sequent 'VrVyVz((Lry 6c
Lyz\ - Lxz), Lab, Lbc F Lac' provides a good illustration (the sequent makes
intuitive sense if you read 'I-l as "is longer than"):
1. VrVyVz((Lxy k Lyz) -Lxz) A
2. Lab A
3. Lbc A
234 Cxaprrn 8
but that would not have been useful with the assumptions given. Use of VE with
multiple quantifiers and multiple names thus requires circumspection and careful
planning.
Our final example, a proof of 'fx-Fx F -VxFx', contains several twists,
requiring in addition to VE the use of EFQ with an 3E strategy that in turn is
nested in an indirect proof:
1. ∃χ∼Fχ A
2。 │ ∀χFχ H(br∼ I)
3. │ │ ∼Fa H(brヨ E)
4.│ │ Fa 2∀ E
5。│ │ P8ζ l∼ P 3,4 EFQ
6。 I P8(∼ P l,3-5ヨ E
7. ∼∀χFχ 2-6∼ I
Exercise 8.3.1
Prove the following sequents:
1. FV*F**Fa
2.∼ Fa卜 ∼∀χFχ
3.∀ χ(Fχ ―→Gχ ),∼ Ga卜 ∼Fa
4.∀ χ(Fχ → Gχ ),Ga→ Ha卜 Fa→ Ha
5。 ∀χ∀γ(Rχ ッ→ Rン ),Rab卜 Rba
6.∀ χ(Fχ → Gχ ),ヨ χFχ トヨχ(Fχ 8ζ Gχ )
7.∀ χ(Fχ ―
→ Gχ ),ヨ χ∼Gχ 卜∃χ∼Fχ
8。 ∀χ(Fχ
―
→ Gχ ),∼ ヨχGχ 卜∼∃χFχ
9.∀ バ Fχ → Gχ ),∀ χ∼Gχ 卜∼∃χFχ
10。 ∀χFχ ∨Vχ G為 ∼Ga卜 ∀χFχ
Crassrcar PRrorcqrr Loctc: lrurrRrnce 235
Exercise 8.3.2
Formalize this argument, using the symbol scheme given below:
Triticale, an artificial cross between wheat and rye, is superior to both.
Since wheat and rye are natural, this proves that it is not the case that every-
thing natural is superior to everything artificial.
One-Place Two-Place Three-Place
Names Predicates Predicate Predicate
Now attempt to prove the resulting sequent. If you symbolized just what was said,
you will find that the sequent cannot be proved; it is in fact invalid. In attempting
the proof, however, you should notice that adding a certain obvious assumption
would make the sequent valid. Identify this assumption, formalize it, and then
prove the resulting sequent.
The last of the four quantifier rules is universal introduction (VI). Like 3E, it makes
use of a representative instance. Nevertheless, it is simpler, since it does not require
a hypothetical derivation. To illustrate it, we shall consider this valid but unsound
argument:
All frogs are green.
Everything is a frog.
Everything is green.
To obtain the conclusion from these premises, we may reason as follows:
Since all frogs are green, if Alice is a frog then Alice is green. And since
everything is a frog, Alice is a frog.It follows that Alice is green. But since
we have made no assumptions about Alice, she is a representative individ-
ual for any object whatsoever. What we have proved of her must be true of
everything. Therefore everything is green.
The last step of this reasoning is a step of VI. Its validity depends upon the fact
that nothing is assumed about Alice; that is, her name does not appear in any
hypothesis whose derivation has not yet ended or in any assumption. This makes
her a representative individual for everything in the domain of discourse. Thus
236 CHAPTER 8
what we prove of her could equally well have been proved of anything-which is
why we can legitimately infer that everything is green.
If we symbolize the argument in the obvious way, we obtain the sequent
'Vr(Fr - Gr), VxFr F VrGx'. Our reasoning may now be formalized as follows:
1. * Gx)
Vx(Fx A
2. YxFx A
3. Fa- Ga 1 VE
4. Fa 2VE
5. Ga 3,4 -E
6. YxGx 5 VI
Step 6 is legitimate only because 'a' does not appear in any assumption or hypoth-
esis so that we could validly have substituted any other name for'a' in steps 3-5.
This shows that we could have proved Gct for any name cr, which is why we may
validly infer'VrGr' at step 6.
Here is the formal statement of the VI rule:
In terms of our example, (D is 'Ga', cr. is 'a', B is 'x', @P/* is 'G.r', and, consequently,
VBOP/" is 'VxGr'. The qualification 'not already in <D' is needed to ensure that we
do not introduce quantifiers of the same variable with overlapping scopes and so
infer something that is not even a formula.
To prove a universally quantified statement, then, we prove a representative
instance of it, making sure that the representative individual is not named in any
hypothesis whose hypothetical derivation has not yet ended or in any assumption.
Then we apply VI. In annotating uses of VI, we cite only the line number of the
representative instance.
With the addition of VI our quantifier rules are complete, and we can now
prove any valid sequent of predicate logic that does not involve identity. Let's
consider another simple example, a proof of the sequent 'Vr(Fx * Gr), Vr(Gx
-
Hx) FVx(Fx*F{rc)':
1; Vx(Fx * Gx) A
2. Yx(Gx * Hx) A
3.1 Fa H (for -I)
4.I Fa-Ga 1VE
s. I Ga 314-E
6.
7.
II Ga-Ha
Ha
2VE
515*E
Cressrcar PRrorcnre Locrc: lrureRrrce 237
8. Fa* Ha 3-7 -l
9. Yx(Fx - Hr) 8 VI
The strategy of the proof is governed by the structure of the conclusion. Here the
conclusion is the universal formula 'Vr(Fx * Hx)', so we must prove a represen-
tative instance of it-that is, a conditional of the form Fct * Ha. To do that we
proceed by conditional proof. We hypothesize 'Fa', taking 'a' as the name of the
representative individual, at line 3. The conditional 'Fa * Ha' is proved at line 8.
Since now at line 9 'a' does not appear in any hypothesis whose hypothetical
derivation has not yet ended or in any assumption, we may apply VI, knowing
that we could have proved Fa .- Hct for any name ct.
The next example combines VI with a reductio strategy. The sequent to be
proved is'-lrFr F Vx-Fr':
1. ∼ヨχFχ A
2.│ │
Fa H(br∼ I)
3。 │ ヨχFχ 2ヨ I
4.│ ∃χFχ ;∼ ヨχFχ l,3&I
5. ∼Fa `文 2-4∼ I
6. ∀χ∼Fχ 5∀ I
∼∀χFχ A
2
∼∃χ∼Fχ H(br∼ I)
3
∼Fa H(fOr∼ I)
4
∃χ∼Fχ 3∃ I
5
∼∼Fa 3-5∼ I
7
Fa 6∼ E
8
∀χFχ 7∀ I
9
There is no obvious way to proceed once the assumption has been set out, so we
hypothesize the negation of the desired conclusion for indirect proof at line 2.
Now we notice that we could contradict assumption 1 if we could prove 'Fa' and
then get 'VrFx' by VL To prove 'Fa' we work once again by indirect proof,
238 Cunpren 8
hypothesizing '-Fa' at line 3. From there the rest of the proof plays itself out as
planned.
The next example illustrates once again the utility of EFQ within quantifier
proofs. It is a proof of the sequent'-1xFx F Vr(Fr * Gx)':
1. ∼∃χFχ A
2. I Fa H(fOr→ I)
3. │ ∃χFχ 2∃ I
4.I Ga l,3 EFQ
5. Fa― →Ga 2-4-→ I
6。 ∀χ(Fχ ―→Gχ )
5∀ I
We could have avoided the use of EFQ by hypothesizing '-Ga' for -I at line 3, but
this would have made the proof several steps longer. This sequent is of intrinsic
interest, since it says that from the premise that nothing is F we may infer that all
F are G. Here we see once again (recall the discussion in Section 7.3) the odd
behavior of universally quantified material conditionals whose antecedent terms
are empty.
As in propositional logic, the general strategy for proving a sequent in pred-
icate logic is determined by the form of the conclusion or subconclusion for which
we are aiming. The proof strategies in Table 8.1 supplement those given in Table
4.1 in Section 4.3 and should be used in combination with them. Thus, for exam-
ple, if the conclusion to be proved is of the form'VrFr
- YxGx', which is a
conditional, first apply the strategy for conditionals from Table 4.1. Thus we
hypothesize'YxFx'and aim to derive the subconclusion'VxGr'. This subconclu-
sion is of the form VB<D so that we next apply the strategy for that form from
Table 8.1. Using the two strategy tables together in this way, most problems can
be solved fairly easily. However, some proofs in predicate logic demand ingenuity
that cannot be encoded in a simple set of instructions.
TABLE 8.1
Proof Strategies
'x' andO is the open sentence 'Fr'. For given such a proof, it is easily seen that any
instance of the corresponding rule may be proved by precisely the same sequence
of steps-that is, by a proof of precisely the same form. Take the rule 'From -VB@,
infer 3B-@', for example. We proved the corresponding sequent'-VxFx llx-Fx'
above. But by precisely the same steps we could have proved '-Vx(Fx & Gr) l-
3r-(Fr 6c Gr)' or '-VyGy l1y-Gy' or '-YylxLxy l1y-lxLxy'-ot any other
sequent having the form indicated by the rule. To show this, we shall rewrite the
proof of '-YxFx l1x-Fx' in full generality, using the Greek metavariables B and
<D in place of the variable 'r' and the open sentence 'Fx', respectively. It is evident
by inspection of this proof form that no matter which variable we use for p or
which open sentence with free variable B we use for (D, the resulting sequent can
still be proved using only the basic rules:
1
∼∀βΦ A
2
-lg-o H(br∼ 1)
3
∼Φα
/β H(br∼ I)
4
ヨβ∼Φ 3∃ I
5
ヨβ∼Φ退に∼∃β∼Φ 2,4&:I
6
∼∼Φり 3-5∼ I
7
Φα 6∼ E
│
/β
240 Cxnpren 8
8。 │ ∀3Φ 7∀ I
9。 │ ∀
bΦ &∼ ∀
βΦ l,8&I
10。 ∼∼∃β∼Φ 2-9∼ I
ll・ ヨβ∼Φ 10∼ E
Here @"/u is a formula that results from replacing each occurrence of B in the open
sentence @ by some name ct not already in <D. From this proof form it is clear that
'From -VPO, infer 3B-<D' is a derived rule.
In the same way, we can see for each of the other quantifier exchange rules
that by proving its corresponding sequent, we have in effect shown that every
instance of that rule can be proved using only the fourteen basic rules ofpredicate
logic. That is, we can see that the quantifier exchange rules are in fact derived
rules.
Exercise 8.4.1
In the text we verified that three of the eight quantifier exchange rules are derived
rules by proving the sequents '1x-Fx l- -VrFr', '-lxFx I Yx-Fx', and ,-VrFr F
1x-Fx'. Verify this for the other five quantifier exchange rules by proving the
following sequents using only the fourteen basic rules:
1. VrFr I -lx-Fx
2. lxFx F -Vr-Fr
3. Vr-Fr I -1xEx
4. -Yx-FxllxFx
5. -3r-Fr F VxFx
Exercise 8.4.2
Prove the following sequenrs, using basic or derived rules:
1.∀ Ц Fχ ―
→ Gχ )卜 ∀χ(∼ Gχ → ∼Fχ )
2。 ∀バ Fχ → Gχ )卜 ∀χFχ → ∀χGχ
3。 ∀χ →
(Fχ ― Gχ )卜 ∀χ∼(Fχ 8ζ ∼Gχ )
5。 ∀χ(Fχ ―
→ Gχ ),∼ ∀χGχ 卜∼∀χFχ
6.∀ χ(Fχ → G→ ,∀ χ∼Gχ 卜∀χ∼Fχ
7.∀ χ(Fχ ―
→ Gχ ),∼ ヨχGχ 卜∀χ―Fχ
8。 ∀χ →
(Fχ ― Gχ ),∼ ∀χGχ トヨχ∼Fχ
9。 ∀χRχ χ卜∀χヨγRχ ツ
10。 ∃ χ∀ッRχ ット∀ッヨχRχ ツ
11.卜 ∃χ(Fχ &∼ Gχ )― ∼∀χ(Fル → Gχ )
′
12. トヨχFχ → ∼` ▽χ∼Fχ
13. 卜∀χFχ ←)∼ ヨχ∼Fχ
14. 卜∀Ц Fχ → Fχ )
Exercise 8.4.3
Prove the sequents corresponding to the arguments formalized in Exercise 6.1,
using basic or derived rules.
8.5:DENTITY
To prove sequents involving identitS we need two more rules, =l and =E. The
identity introduction rule is utterly simple:
In a sense, this is not even a rule of inference, since it employs no premise' It simply
licenses us to assert a logical truth of the form o = o at any line of a proof. Clearly
this is a valid procedure since a logical truth must be true given any premises-or
none at all. For annotation, we simply write '=I' to the right. Since no premise is
used, no line number is listed. This proof of the sequent 'VrVy(r = y'Rxy) f
Raa' illustrates a simple use of =I:
l. YxYY(x=Y-R Y) A
2. YY(a=Y*RaY) 1VE
3. a=a*Raa 2VE
4. a=a =I
5. Raa 3,4'E
The following proof of the theore m'l -lx-x = r' exemplifies another typical
use of =I:
1
1x-x = x H(fOr∼ I)
2
∼a=a H(forヨ E)
3
a:=a =I
4
P8ζ l∼ P l,2-4ヨ E
6
-1x-x = x l-5∼ I
Uses of =E are annotated by citing two line numbers: the number of the line on
which q = B occurs and the number of the line on which @ occurs.
The validity of =E is fairly obvious, but to prove rigorously that it is valid
takes considerable work, and we will postpone that task until Section 9.1. In the
meantime, we shall simply use it. Here it is used in a simple proof of the theorem
'F(Fa&a=b) *Fb':
1
Fa 1&E
3
a=b 18こ E
4
Fb 2,3=E
5
Exercise 8.5.1
Prove the following sequents:
b メ 〓 νン じ
,
F ∀
ー
F ト
に χ a χ
a ¨
a,
D ↓
” 〓
跛
一
一一
,
R
↓
2
χ
γ
y
χ χ χ χ
●
′
.
ヨ ∀ ∀
3
χ ﹁ コ 4ν
4. F
νン
〓 χ
一
D
ⅣⅣⅣ掏W
一
けつつ 一
月 & 陣 .ブ
方烈 〓︺γ
〓 y ﹁り
↓ χ ν′“ ↓
6
〓∀ ● a じ
χ χ χ しい ︼
νン 一
ν 片,F
ハ´
︱ も
↓D J l
χ χ 卜月
z I R
7
〓 ↓ ”
V
一
κ T^
一
一 一 ・
ツヨ
8
aI
・ L
・ガ
︰ r
,に
9
卜
・ ,
ヽ
l
b
,
﹁
0
a
▽
″ ‘
′
Exercise 8.5.2
Prove the sequents corresponding to the arguments that were formalized in Exer-
cise 5.3.
8.6 FUNCT10NS
1. I a=b H(fOr→ I)
2. I f{a1=61r, =I
3. I fral= f16, l,2=E
4. a=b-f(a) =f(b) l-3→ I
5. VY(a =Y-f1a1=f1Y1) 4∀ I
6. YxYY(x = Y - f1Y1 = f1Y7) 5∀ I
Note the use of =I with the complex name 'f(a)' at line 2 and the quantification of
simple names contained in complex names at steps 5 and 6.
2M Cunprrn 8
The second of our two theorems, 'F Vx3yVz (7 = frr1 * z = !)' , says that for
each argument (represented here by the variable 'x') a function has a unique value
(represented by'y');
1. I b=f(a) H(for-I)
2. b=f1ay*[=6111 1-1-I
3. b=f1ay-g=f121 2,2*I
4. Yz(z=f(a)*z=f\a)) 3VI
5. lyYz(z=f1a1*7=y) 4=l
6. YxlyYz(z=f@) * z=!) 5 VI
Of special interest in this example is the use of lI with the complex name 'f1a;' at
line 5.
Exercise 8.6
Prove the sequents corresponding to the arguments formalized in Exercise 5.4.
c〃 ハ PrfR
Recall that a test for validity is sound if whenever the test classifies a sequent as
valid it is valid. A test is complete if (conversely) whenever a sequent is valid, the
test classifies it as valid. Thus to say that a test is sound and complete is to say
(1) The test classifies a sequent as valid if and only if it is valid.
But if this is the case, it would seem that the test is, after all, a decision proce-
dure-for wouldn't it have to give a correct answer in every case in order to be
sound and complete?
Actually not. For statement (1) does not tell us exactly what will happen if
the sequent is invalid. tilfhat it does tell us (by contraposition) is this:
(2) A sequent is invalid if and only if it is not the case that the test classifies
it as valid.
Now there are two ways that the tree test might fail to classify a sequent as valid:
(i) the test might classify rhe sequenr as invalid; that is, the rree might ter-
minate with an open path, or
(ii) the test might not classify the sequent at all; it might go on forever and
never return an answer.
245
246 Cuaprun 9
The following lemma, then, demonsrrates the extensionality of all formulas. (In-
cidentall5 this lemma illustrates in great detail the straregy of arguing by cases-a
strategy analogous to vE in the object language.)
‐│111111111111111‐ │'│■ │
'We
.
γ(中 )(Ψ α )
T iff γ(m(β )8γ (β ,′
ソ β
(α .)(Φ )=γ (田 (Φ /α )
│‐
菫ёncc ‐ ‐
‐ ││‐ ‐
‐
‐ ‐ │││‐ │‐ │││‐
‐
・
).=Ti“ くγ プ)(δ l),...ち γ 轟〕
‐ γ(機 )(Ψ 、1,..:‐ ,γ ″ (δ ″ )>│││‐ lε
‐ ‐
(β
_‐ _ ‐
‐7(凛 ,(Ψ γll...:lγ ″)I T iff<γ ′ ・・,V(β (嵐 )(δ l),・ ).(0″ )≧ │││
l ‐
,ン
‐ γ ,7)(Ψ ) ‐ ‐ ││‐
(β . ‐
‐‐ │
‐ ││ ││‐ │‐
1‐ lence by bivalcnce
. γ(中 )(Φ )=γ (β ,.)(Φ
/.).
β ‐
│
except that it uses valuation rule 3 rather than valuation rule 2.) This
││
「 │
establishes that all formulas of complexity 0 are extensional.
1 )st set in S
.lS,
show that all formulas of
vve lllust +111■ rel lll
extensional. Now
plexity z + 1 must have one of the following formsr
negatlon biconditional
con,unctlon
disjunction
'We
now prove a second extensionality lemma, which is a simple corollary of
the first lemma:
Incidentally, the validity of the inference rule =E, which we did not have the
means to confirm in Section 8.5, follows directly from the special extensionality
lemma. The proof is left as an exercise.
Having established these lemmas, we are now ready to prove the soundness
of the tree test. As in Section 5.8, we prove first that if the initial list is consistent,
there is an open path through any tree obtainable from it, and from this the
soundness result follows as a corollary.
│││‐ ‐‐
││││
│‐
of the other rules. Now these rules, when applied to a path all
of whose formulas are true on a valuation, yield at least one
one-step extension all of whose formulas are true either on
that valuation or (in the case of f V when used to intro-
-and
duce a new name) on an expansion of that valuation with
respect to the new name.2 (Check this claim for each rule. The
'z This is the case for ! and V, provided that the original valuation V' did not assign
any extension to the new name; that is, it is a potential name for each obiect in the
domain with respect to Y'. But these rules require that the new name not appear
on the path P, and our inductive hypothesis ensures that V' assigns no extensions
to names not on P.
254 Craprrn 9
Soundness now follows by essentially rhe same reasoning used in Section 5.8:
3 Suppose, for example, that I is applied to some formula 3B@ to yield a new for-
mula @r/p, which is the result of replacing each occurrence of B in O by some new
name "y. Now, if there is some valuation V with domain I such that Y(3pO) = I
then by valuation rule 10 there is some potential name o of some object I in I
such that "l,r,"..,,(O"/u) = T Now since 1 is a new name, it is not assigned any value
by V, and so it is also a potential name for L And since Y,r.,,(t) = Yr", r(cr)= ;, it
follows by the general extensionality lemma that 1/,,,, ,(@,/u) = V,",, r(<D"/B) = T
Therefore there is some expansion of 1r-namely, Y,r',-that makes (Dr/u true.
Cmsstcar Pneotcarr Loctc: SoutrtoNess, CovprrteruEss, AND lNexpResstgtLtrv 255
‐
│■ ││‐ . . . ; @,, -Y, and that all paths of the tree close after
│● 11■
Exercise 9.1
1. In the basis case of the proof of the general extensionality lemma, we left
unproved the assertion that all identity statements (statements of the form ct =
B) are extensional. Prove this.
2. In the inductive step of the proof of the general extensionality lemma, we
showed, given the hypothesis that all formulas of complexity n or less are
extensional, that all coniunctions of complexity n + 1 are extensional. Com-
plete the proof of this lemma by showing, given this hypothesis, that all nega-
tions, disjunctions, conditionals, biconditionals, existentially quantified for-
mulas, and universally quantified formulas of complexity n + 1 are also
extensional.
3. In the proof of the main metatheorem of this section, it is necessary to verify
for each tree rule other than - or -= that when applied to a path all of whose
formulas are true on some valuation "l/, it yields at least one one-step extension
of that path whose formulas are all true on V or on some expansion of V. The
propositional rules were covered in Section 5.8, and I did the checking for I in
footnote 3 of Chapter 9. Complete the necessary verifications for -V, -f, V,
and =.
4. Prove, using the special extensionality lemma, that all instances of the inference
rule =E are valid.
A test for validity is complete if whenever an argument is valid, the test classifies it
as valid. In this section we prove the completeness of the tree test for predicate
logic with identity but no function symbols. As usual, we first prove a metatheo-
rem and then derive the completeness result as a corollary. The metatheorem as-
serts that if a tree never closes, its initial list is consistent. To prove this, we use a
255 Crnprrn 9
'we
strategy like that of Section 5.9. begin by supposing for conditionalproof that
we have a tree that never closes. In propositional logic, where all trees are finished
after a finite number of steps, this would simply mean that the finished tree con-
tains an open path; but in predicate logic it is also possible that the tree fails to
close because it is nonterminating, that is, infinite. Yet if it is nonterminating, we
know by Konig's lemma (lemma 3, Section 5.7) that ir contains an infinitely pro-
longable path. Hence (whether or not the tree is infinite) if it never closes, it
contains a specific path that never closes. We saw in a practical way in Section 7.4
that such a path, whether infinite or not, defines a valuation Y which makes each
of its members, including the entire initial list, true.4 But the existence of such a
valuation establishes (by the definition of consistency) that the initial list is consis-
tent, thus completing the conditionalproof.
In outline, then, the proof is straightforward. The real work comes in defin-
ing Y and proving that it really does make each member of the open path true.
Here are the details:
METATI{EOREM: If a tree
with identity (but no
ls conslstcnt.
nating or it is not. If L
produces an infinitely path. If L
nating, then L yields a
cations of the rules; and
at least one open path. Hence in either
which never closes. We now
hence
a set of
name 'a' denotes the number 1,
so on! except that where names identity │・ ■│■ │
│● ││
.│‐
││‐
││.││ ││ .‐
if @ is an n-placepredicate (n> 1),then for any numbers
lr., . . ., /,in,b, 1/r, . . ., /,,> e1/'('@') iff there are names
.│‐ ││.‐
1‐ │‐
.││.│‐
‐ ‐
││││.
│ │‐
or, . . ., cr, such that for each i, 1. < i < n,Y(a,) = Ji, a.nd
@cr, . , . ) at appears as a line of P.
5 The identity rule might also be applied to formulas of any of these forms, but
whether or not it is, still one of the rules listed must eventually be applied.
258 Cuapren 9
Hence P will
on ? on some path,
on Y, since
.l/ is constructed
―
ICOROLLARY (Completeness): If a Seq,Cnt Of PrediCate loJこ ‐ ■ith
identiry (but no function symbols) is validぅ the trё e tcst。 lassines that
sequent as valid. │││││. │■ │││││
Set{Φ l,...,Φ
",
it is not the
eventually does close. But this is to say that the tree testclassi-
fies the sequentO,, . . ., O, F Y as valid.
Crassrcal PRrotcare Locrc: Soutorurss, CovprrrrnEss, AND lNexpnrssrsrrrrv 259
\7e have now shown that the tree test is both sound and complete for predi-
cate logic with identity but no function symbols; that is, it classifies a sequent of
this logic as valid if and only if that sequent ls valid. Conversely, this means that
the tree test fails to classify a sequent as valid if and only if it is not valid. But
failing to classify a sequent as valid is not the same thing as classifying it as invalid.
Nonterminating trees fail to give an answer either way, and we know that such
trees exist. So, though we are guaranteed a right answer if we apply the test to a
valid sequent, and though we are guaranteed that if we apply the test to an invalid
sequent and get an answer that answer will also be right, there is no guarantee that
we will get any answer at all when we apply the test to an invalid sequent. This
means, further, that if we apply the test to a sequent which we neither know to be
valid nor know to be invalid, we can't be sure in advance that we will learn
anything. The tree test for predicate logic, then, is not a decision procedure.
This raises the question of whether we might modify the tree test somehow
to make it a decision procedure-or invent some wholly different procedure to
decide the question of validity in predicate logic. The answer to this question is
no. To prove that, we will need to say more precisely what a decision procedure is.
We have defined a decision procedure as a terminating algorithm which determines
for each sequent of a given logic whether or not that sequent is valid. The term
in this definition which needs clarification is 'algorithm'. We shall clarify it in
Chapter 10.
TABLE 9.1
Tree Rules and Corresponding Proof Steps
Universal Quantification (V) If a formula In either case, infer @"/u from VB<D
of the form VB@ appears on an open path by VE.
and ct is a name that occurs in a formula
on that path, write <D"/u (the result of
replacing all occurrences of B in @ by ct)
at the bottom of the path. If no formula
containing a name appears on the path,
then choose some name ct and write @'/o at
the bottom of the path. In either case, do
zot check VB<D.
Negated Identity (-=) Close any open Introduce o. = a by =I, then from ct =
a formula of the form -ct = ct and -a = ct, deduce'P & -P' by
l"j:;lJn'.n EFQ.
Crassrcal PRgotcaru Loctc: Souruottrss, Couprete NE55, AND lNrxpResstatLttv 261
The soundness proof for the inference rules of predicate logic is also the same
as the soundness proof in Section 5.10, except that the basis case now must take
account of =I, and each of the six new inference rules must be checked in the
induction to ensure that they yield valid corresponding sequents. The proofs are
relatively straightforward and are left as an exercise. We thus are able to show that
the inference rules for predicate logic are sound.
To summarize: The inference rules for predicate logic are both sound and
complete. That is, a sequent is valid iff it can be proved. But we have been unable
to foimulate, as we did for propositional logic, a terminating algorithm for telling
which sequents are valid and which are not. As a result, when given a sequent of
predicate logic, we cannot always tell whether or not it is provable. In Chapter 10,
we shall see that no terminating algorithm of any kind can decide that question in
every case. In the meantime, however, we consider another reason we cannot
always tell whether an argument is valid.
Exercise 9.3
1. Write out in full the completeness proof for the inference rules of predicate
logic, using the completeness proof of Section 5.10 as a model.
2. Show that the soundness proof of Section 5.10 may still be applied to predicate
logic by proving for each of the rules fI, fE, VI, VE, =I, and =E that when
applied to lines whose corresponding sequents are valid, they yield a conclusion
whose corresponding sequent is valid. (Hint: Some rules require the extension-
ality results of Section 9.1.)
9.4 INEXPRESSIBILITY
Predicate logic does not automatically decide the validity of every argument. Even
if we manage to formalize an argument satisfactorily, there is no guarantee that
we will be able to decide the validity of the corresponding sequent. But sometimes
we cannot even get that far. In these cases the problem may lie in the inexpressibil-
ity in predicate logic of crucial elements of the argument's form.
ttr. fact that some forms of reasoning are not expressible in predicate logic
is hardly surprising. Predicate logic deals with a mere eight operators: '-','&-','Y',
'+','e', 'V', 'l', and (-'-16gether with the operators expressible by combining
these eight.6 But why suppose that these are the only expressions whose semantics
can affict validity? The validity of an argument form could conceivably depend,
for example, on the semantics of certain predicates, such as 'is a part of' or 'is
greater than'; adverbs, such as 'quickly' or'necessarily'; or sentence operators,
i.rch as '. . . knows that . . .' or 'it should be the case that . . .'. In such cases,
predicate logic might misjudge validity, for it takes no account of the semantics of
such expressions.
A natural response to the problem of inexpressibility is to expand predicate
logic to include more and more of these "nonstandard" operators. Part IV of this
book explores that response. Not all logicians endorse it. 'Nonstandard' is an
epithet used derisively by some, who argue that some or all of the operators we
examine in Part IV are not really logical, or not really intelligible, or both, and
that logic proper ends with predicate logic. This, however, is not my view.
In this section we consider some arguments whose validity is not decided, or
seems to be incorrectly decided, by predicate logic. In each case we shall see that
the problem originates with one or more expressions whose semantics is not fully
representable in predicate logic. Then we consider ways of expanding predicate
logic to accommodate the semantics of some of these terms. our first example is
This finger is part of this hand.
This hand is part of my body.
This finger is part of my body.
Using 'f' for 'this finger', 'h' for 'this hand', 'b' for 'my body', and 'p' for .is a part
of', this argument may be formalized as
Pfh, Phb F Pfb
But while the argument is intuitively valid, rhe sequent is invalid. !7hat has gone
wrong?
The stock response is that this argument has a missing premise, namely, that
if one thing is a part of a second and the second a part of a third, then the first is a
part of the third. In symbols this is
VrVyVz((Pry k Pyz)
- Pxz)
Or we might add merely the relevant instance of this generalization:
(Pfh & Phbl-- p66
Adding either premise does indeed make the sequent valid in predicate logic, but
it evades the question. we wanted to know whether the original argument had a
valid form-not whether the argument that results from adding either premise
does.
The point can be illuminated by a somewhat different example. Suppose
someone were to claim that the argument
If all men are mortal and Socrates is a man, then Socrates is mortal.
so that the argument really is
Cussrcnl PRrorcarr Looc: Souruorurss, CovprererEss, AND lNrxpRrsstatltrv 253
If all men are mortal and Socrates is a man, then Socrates is mortal.
All men are mortal.
Socrates is a man.
Socrates is mortal.
This is clearly an arbitrary addition. Why stop here?'We could as well argue that
this augmented argument is still incomplete, since it in turn "really assumes" the
further premise
If it is the case that if all men are mortal and Socrates is a man, then Socra-
tes is mortal, and in addition all men are mortal and Socrates is a man; then
Socrates is mortal.
We have started on an infinite regress, so we had better call a halt. The only
nonarbitrary stopping place is the original argument. It was valid as it stood.T
AnalogouslS the argument
is arguably valid as it stands. Why do some people want to add a premise to this
argument, even though they would not be tempted to do so with the Socrates
argument? The answer, I think, is that if they didn't postulate the hidden assump-
tion, they would have to admit that predicate logic, which pronounces the argu-
ment's form invalid, gives the wrong result. The addition is made ad hoc, to
salvage agreement between formalism and intuition. As a result, it is not the for-
malism that decides the argument's validity. The question of validity is decided by
intuition before the decision to add or not to add a premise is made; then, since
the form is seen to be valid, a premise is added to square predicate logic with
intuition.
In this way, any argument that seems intuitively valid can be made formally
valid by the addition of premises. But so can any argument that doesn't seem
intuitively valid, even those that are egregiously invalid. There are many ways to
make such arguments valid. One that always works is conditionalization. Given
any sequent @,, . . . , O, F Y, simply conjoin the premises and make their coniunc-
tion the antecedent of a conditional whose consequent is the conclusion. Adding
this conditional as a premise yields the valid sequent:
((@,&. . . &O")*Y),@,, . . ., O, l- Y8
7 Lewis Carroll, who was a logician as well as a writer of children's literature, dis-
cusses this regress of premises enlighteningly and entertainingly in a well-known
article titled "lVhat the Tortoise Said to Achilles ," Mind 4 (1895 ): 278-80.
8 I have suppressed brackets in the conditional's antecedent. They must be added to
make the conditional a formula, but all ways of adding them are equivalent so
that their exact placement does not matter.
264 Cnrprrn 9
So, for example, even such classical fallacies as the following ad bominem become
valid in virtue of their "hidden" premises:
Bill believes that God exists.
Bill is an idiot.
God does not exist.
If Bill believes that God exists and Bill is an idiot, then God does not exist.
Nfith so flexible a procedure, it's no wonder that predicate logic can be made to
yield correct results.
But because it can transf orm any argument into a valid argument, this pro-
cedure will err unless the decision to add or not to add premises is based on an
accurate-intuitive assessment of the validity of the argument before the addition of
premises. It is informal intuition, not the formalism, that decides initially whether
or not the argument is valid and thus guides our decision to add or not to add
premises.
This is not to deny the general usefulness of adding premises. On the con-
trary, making unstated premises explicit is often essential to accurate argument
analysis.e The point is only that it is possible to evaluate arguments intuitively
before adding premises, as well as afterward, and that in many cases when we
evaluate arguments before adding premises we seem to find them valid, though
predicate logic pronounces their forms invalid. We can always compensate for
e But it is essential only where the suppressed premise is not logically true. The
addition of a logically true premise does not alter the validity of an argument and
so is merely redundant. That is why it is pointless to add a premise like 'If all men
are mortal and Socrates is a man, then Socrates is mortal'. Likewise, the principle
'If one thing is a part of a second and the second a part of a third, then the first is a
part of the third' is arguably a logical truth (by the semantics of the predicate 'is a
part of'), though predicate logic takes no account of this. It would follow that
adding this premise, too, is pointless, except to make predicate logic square with
what we already know by intuition.
The addition of a premise that is not a logical truth is, by contrast, not
redundant. In the argument
Jill is human.
.'. Jill is fallible.
this, because we can always transform an invalid argument into a valid one by
adding a premise. But still it seems that predicate logic incorrectly classifies the
original.
'u7ith
Why? our example, specifically, the argument's validity seems to hinge
on the meaning (i.e., the semantics) of is a part of'. Except for the identity predi-
cate, predicate logic makes no provision for representing the semantics of specific
predicates. In fact, it systematically ignores their meanings. The sequent
which represents our argument could as well be used to symbolize the argument
Fred praised Hannah.
Hannah praised Bill.
Fred praised Bill.
or any of an infinite number of other arguments, both valid and invalid, with
wildly unrelated meanings. The meanings of individual predicates (other than the
identity predicate) are lost in the process of formalization. Thus it can be argued
that the appropriate form and the relevant semantics for this argument are inex-
pressible in predicate logic, just as the appropriate forms and the relevant seman-
tics for quantificational arguments are inexpressible in propositional logic. W'here
validity depends on the semantics of terms other than the eight operators men-
tioned earlier, predicate logic is insensitive to it.10
In fact, predicate logic seems incapable of symbolizing whole categories of
terms. Consider adverbs. Predicates are at least syntactically representable in pred-
icate logic, though formalization strips away their meaning; but we don't even
have the syntax to symbolize adverbs. In the argument
Qb r- \7b
which is invalid. Yet the argument seems valid.
There is, however, a way of conceptualizing the argument's form within
predicate logic that just might work. Donald Davidson has devised an ingenious
analysis in which adverbs, prepositional phrases, predicates, and several other
1e((De&Teb)&He)
where 'D' stands for'driving', 'T'for'through','b'for'Boston', and'FI'for'was
harrowing'. The same formalization could be used for 'It was harrowing to drive
through Boston', which uses the infinitive 'to drive through Boston' instead of
the gerund.
Thus Davidson's analysis is widely applicable. But it succeeds only at the
price of awkwardly complicating the syntax of predicates and of treating events as
objects. As a result, it has its critics.
Yet this analysis stands for us as a caution. Because predicate logic contains
no explicit symbols for adverbs and prepositions, it would be easy to conclude
that it cannot deal with them. This, as Davidson reminds us, does not follow. The
expressive capacity of predicate logic, intelligently deployed, may surprise us.
So far we have considered arguments whose intuitive validity seems obvious;
the problem is that intuition doesn't agree with predicate logic. Whether our next
and final example is valid is zol intuitively clear. Suppose Bad Bart, who is at this
very moment oozing in through the window, is bent on murdering me and can be
stopped only by being killed. Then (if I have time) I may reason as follows:
I should live.
It is necessarily the case that if I live Bad Bart dies.
Bad Bart should die.
Exercise 9.4
Each of the following arguments is, or might reasonably be thought to be, intui-
tively valid. Try to express the form of each in predicate logic. If you fail to produce
a valid sequent, discuss the reasons for your failure. Don't add any premises!
10.l ABACI
In this section we discuss the design and operation of a rudimentary sort of com-
puter called an abacus (plural abaci), or register machine. Our goal is to analyze
the notion of an algorithm, which, as we shall see, turns out to be definable as
an abacus program. Having understood this concept, we will be in a position to
prove that predicate logic is undecidable-that is, it is in principle impossible to
devise an algorithm that tells for any sequent of predicate logic whether or not it
is valid.
Abaci should not be confused with the ancient counting devices of the same
name, though they have some similar features. An abacus can be thought of as a
series of bins, called registers, together with a control mechanism for moving
counters (which you might think of as stones or marbles) one at a time into and
out of these bins. (See Figure 10.1.) Programming an abacus means telling this
mechanism when and where to add or remove a counter. ln tbeory, the abacus has
an unlimited number of registers and an unlimited supply of counters, and each
I For the material in this chapter, I am indebted to Richard Jeffrey's elegant treat-
ment in Formal Logic: Its Scope and Limfts, 3rd ed. (New York: McGraw-Hill,
1991), Chaps.7 and 8.
268
Ct.essrcar PRe otcarr Loctc: Ut{orctoastltrv 269
F:GURE 10.1
inlaginative Representation of an Abacus(Register Machine)
Control
Mechanism
│ ││ ││ ││ ││ │
│ ││ ││ ││ ││ │
│ ││ ││ ││ ││ │
担 L_」 L」 L_」 L_」 ヒ _」
lnfinite pile Register A Register B Register C Register D Register E
of extra
counters
register can hold an unlimited number of counters. But in any real machine these
quantities must be finite.
Abaci are simpler in design than ordinary computers, but they work in much
the same way. Ordinary computers have a large number of registers, each of which
is capable of storing a small number of zeros and ones, coded electronically. The
list of zeros and ones in a register can be thought of as a number in binary code,
and this number in turn may be regarded as representing the number of "counters"
in the register. Programming a computer amounts to instructing its "control mech-
anism" (central processing unit) to move and manipulate these "counters," as in
an abacus; but the instructions may be more sophisticated than iust "add one" or
"subtract one," which are the only commands an abacus "understands."
Yet this seemingly important difference between computers and abaci is ac-
tually rather superficial; for by giving the abacus sufficiently complex concatena-
tions of "add one" and "subtract one" commands, we could make it mimic any of
the operations of an ordinary computer. It would, however, be very slow, since it
would have to perform many separate additions and subtractions to mimic a single
operation of an ordinary computer.
Not only can an abacus do anything any actual computer can do; there is
good reason to believe that it could do anything that any possible computer could
do (this claim is called Church's thesis, which is the topic of Section 10.4). What is
interesting, then, is to find out what an abacus can't do, since that sets limits on
the capability of any possible computer.
But let's first see what an abacus can do. Suppose we want to program an
abacus to add two numbers. For this we need two registers, register A and register
B, to store the numbers. To add numbers x and y, we put ,c counters in register A
and y counters in register B and instruct the machine to begin taking counters out
of register B and to put a counter in register A for each one it removes from register
B. The process is completed when register B becomes empty. Register A then
contains r + y counters, the sum ofx and y.
270 Crepren 10
FIGURE 10.2
The Two Operations That an Abacus Can Perform
Putting the counters in the registers initially is not part of programming the
machine. The program is merely a general list of instructions for crearing a sum.
!7hat registers A and B contain to begin with is called the input. Once a program
is written, we provide different inputs each time we run it. To progiam an abacus
to do a particular task, we must analyze that task into an ordered series of simple
operations. Fundamentally, an abacus can perform only two operations:
It is easy to see that this set of instructions will result in the adding of the number
in register B to the number in register A, no matter how many counters the two
registers contain initially.
'We
can represent these instructions more perspicuously in a flow chart made
up of addition and subtraction nodes:
一 ●一 ・
.
A+ ■
│・ │・
・
‐ .
一 一一
2
^
U
In this flow chart, the node for register A is an addition node and the node for
register B is a subtraction node. The arrows are numbered for easy reference. (This
is not, however, merely a convenience; it will be crucial later, when we describe the
program in logical notation.) One and only one arrow leads into a flow chart; it is
called the entry arrow. The entry arrow must always be labeled as arrow 0, as it is
here, since the mechanism that operates the program always looks for the 0 arrow
to begin the program. The numbering of the remaining arrows is arbitrary, but
usually there is a more or less natural way to do it. Any arrow which, like arrow
3, leads out of the flow chart (i.e., which does not lead to further instructions) is
called an exit arrow. Exit arrows are the program's stopping points. Not all abacus
programs have them.
The operation of an abacus program can be thought of as taking place in a
discrete series of numbered times. At each time, the machine is following an arrow
in the program. Operations are performed between one time and the next. The
program always begins with the machine on the entry arrow at time 0. To see how
to read the flow chart, suppose we start the machine with three counters in register
B and one in register A. Now at time 0 the machine is on arrow 0. Arrow 0 directs
272 CHAPTER 1 0
it to the node for register B (the bottom node), which gives it the following
rnstructron:
Since register B is not empty, the machine removes a counter, leaving two counters
in register B, and moves on to arrow 1. It is on arrow 1 at time 1. Arrow 1 leads
the machine to the node for regisrer A (top node), where it obeys the instruction
to add a counter to register A and move on to arrow 2. Thus at time 2 the machine
is on arrow 2 with two counrers in register B and two in register A. Arrow 2 leads
the machine back to the bottom node. Since register B is still not empty, the ma-
chine again follows the instruction ro remove a counter from register B and move
on to arrow 1. So at time 3 it is on arrow 1 with one counter in register B and two
in register A. It moves to the top node, adds a counter to register A, and passes on
to arrow 2 again. Thus at time 4 the machine is on arrow 2 with one counter in
register B and three in register A. It thus returns once again to the bottom node.
Since register B still contains a counter, the machine removes this counter and
moves to arrow 1. So at time 5 the machine is on arrow 1 with register B empty
and three counters in register A. By time 6 it is on arrow 2 again, having added a
counter to register A. Register A now contains four counters and register B is
empty so that when the machine gets back to the bottom node, it passes on to
arrow 3 instead of going back to arrow 1. Thus at time 7 the machine is on arrow
3, and, having no more instructions to follow, it halts. Register A now contains
four counters, the sum of the initial values of registers A and B. The same program
would likewise add any other two inputs for the rwo registers.
This predicate thus allows us to describe the state of the machine at a given time.
Each of the four places stands for a positive whole number. The first place follow-
ing the predicate is always the time, the second always stands for the number of
the arrow that the machine is following at that time, and the remaining places
stand for the numbers of counters in the registers, the registers being listed in
alphabetical order. That is, the third place stands for the number of counters in
register A, the fourth place for the number of counters in register B, and so on. So,
for example, the statement
R3 1.27
says:
Arabic numerals such as '0'r'1' , or '1'47' are, grammatically speaking, just names.
IiTe could introduce them as new names into the predicate calculus, or we could
think of them as abbreviations for symbols constructed from the numeral 0 and
the successor sign. If we chose the latter course, we would write the number a as
the zero symbol followed by a successor symbols (symbolized as apostrophes).
Thus the numeral '5', for example, would be an abbreviation for
0rilrr
use the successor symbol primarily with variables.'We'll generally write nu-
\tr7e'11
A unchanged but one less counter (y counters) in register B. This is part of the
instruction set for the bottom node of our flow chart. It is important to notice that
this instruction is not applicable if register B is empty, for then register B could not
contain y', that is, some nonnegative whole number y, plus one, counters. So we
need a separate instruction for the case where the machine is on arrow 0 and finds
register B empty:
Rr0x0*Rt'3x0
This says that if at time / the machine is on arrow 0 with r in register A and register
B emptS the machine should move to arrow 3 without changing either register.
This takes care of the case in which register B is empty when we start the adding
program. In that case, adding register B, that is, zero, to register A will leave
everything unchanged.
Only one more statement is needed to complete the program. 1il7e told the
machine what to do if it is on arrow 2 and finds register B empty, but we haven't
yet told it what to do if it is on arrow 2 and finds something in register B. From
the flow chart, we see that the machine is supposed to take a counter out of register
B and then go on to arrow 1; so we write
P.t2x7'*Rt'1xY
The entire program, then, is
Rf0r/'*Rt'1xY
Rr0x0-Rt'3x0
R/1xY-Rt'2x'Y
P.t2x7'-Rt'LxY
Rt2x0-Rt'3x0
The first two and last two statements of this program together constitute the
instruction set for the bottom node of the flow chart. The third statement is the
sole instruction for the top node.
The reason why the bottom node requires four statements while the top node
requires only one is that the top node has only one arrow (arrow 1) entering it and
only one (arrow 2) leaving, so allwe have ro do is tellthe machine what to do in
going from arrow 1 to arrow 2. But there are two ways to enter the bottom node
(arrows 0 and 2) and two ways to leave it (arrows 1 and 3). '$7hich arrow the
machine takes out of this node depends on whether or not register B is empty.
Thus we need to tell the machine when and how to go from arrow 0 to arrow 1 or
arrow 3 (the first two statements) and when and how to go from arrow 2 to arrow
1 or arrow 3 (the last two statements). So it takes four statements to describe the
bottom node, making a total of five statements for the program as a whole.
More generally, tbe number of program statements needed to describe a node
is the number of arrows coming into tbe node mubiplied by the number of arows
going out. !7hen designing a program, you should always check it for complete-
ness by figuring the number of statements required for each node and then adding
these numbers to get the total number of statements for the program. If your
program contains fewer statements, it's incomplete. If it contains more, you've
written too much.
Crassrcar PRe otcare Loctc: Ut{orctoaatrtrv 275
ABACUS is a computer program that enables you to construct, program, and run
abaci that are simulated electronically within an ordinary computer. Though in
theory an abacus has infinitely many registers, each with infinite capacity, these
simulacra have at most seven registers (you get to choose the number), each of
which holds a maximum of 32,767 counters. Because of the small number and
capacity of their registers, they are not powerful enough to do any serious com-
p,rti.rg,'-rch less to approach the performance of infinite abaci. But they do enable
you to get a feel for how an abacus operates.
Just type
2
You are now ready to enter a program statement. The order in which you
create statements is unimportant (ABACUS will arrange them in the order of their
arrow numbers), but suppose we begin with the first statement of the adding
program, that is,
Rf0x/'.*Rt'1.xY
Now the first place after the R is always the same in all program statements. In the
statement's antecedent it always designates some time t and in the consequent it
designates the next time t'. ABACUS supplies the R predicates, arrows, and time
variables for each statement automatically. Thus what appears on the screen after
you specify the number of registers is
Rt_
Enter initial arrow number
The program is asking for the second place following the R in the antecedent of
the program statement, which in this case is 0. So type '0' and then press either
ENTER or the space bar. The screen now looks like this:
Rt0_
Enter initial value for register A
Register values are designated either by the numeral '0' or by variables (lowercase
letters) with or without a single successor symbol. In the program statement we
are writing, the initial value for register A is designated by the variable 'r'. So type
'x' and then press either the space bar or ENTER. The screen now looks like this:
Rt0x-
Enter initial value for register B
The initial value for register B is 'y". After you enter this, ABACUS supplies the
arroq the second R predicate, and the second time variable in your program
statement, so you see this on the screen:
Rt0xY'-*Rt'-
Enter final arrow number
The "final" settings are those mentioned in the consequent of the program state-
ment. ABACUS has automatically set the final time (the first place following R in
the consequent) to ,' and is asking you for the arrow number, which should be 1.
Enter this and then the final register values'x'and 'y'to complete entry of the
statement. The finished statemenr is now part of your program and is displayed at
the top of the screen.
Correcting Errors
If you make a mistake while entering a statement, you can backspace over it and
retype it, so long as you have not pressed ENTER or the space bar, which enters
data into the program. If you have already entered the mistake, press ESC to abort
the statement and start again.
CrasstcaLPnEorcarr Locrc: Urlorooaerrrrv 277
If you've already entered an entire statment and then found that it contains
a mistake, you can modify it. Use the vertical arrow or the Page Up, Page Down,
Home, or End keys to move the highlighter to the statement you want to modify.
Then press 'm' for MODIFY or use the horizontal arrow keys to move the high-
lighter at the bottom of the screen to MODIFY. Press ENTER. You will see a
window like that used to create the statement originally, except that the statement
to be modified is written at the top. Proceed as if you were creating the statement
'S7hen
anew. you finish typing it, the new version will replace the old in the pro-
gram. If you decide you don't want to modify the statement before you finish
typing, press ESC to abort the modification.
You can also delete a statement without replacing it. Choose the statement
with the highlighter and then choose DELETE at the bottom of the screen and
press ENTER.
If you want to delete all the statements from your program, choose ERASE
from the bottom menu and press ENTER. This erases the program from active
memory but doesn't affect any work that you've saved to a disk.
Saving Programs
To preserve a program once it has been written, you must save it to a disk file;
otherwise, when you start a new program or turn the computer off, the program
will be irretrievably lost. Programs saved to disk files are stored permanently in a
magnetic medium, much as a tape recording is stored on a tape. Once saved, a
program can be called up again into the program editor by using the GET com-
mand. To save a program, choose SAVE from the menu at the bottom of the screen
and press ENTER. You will then see a message that looks like this:
Directory: A:\
Press ENTER to proceed, ESC to abandon, or type a new directory
Unless you are familiar with IBM operating systems and want to use this feature
to save to a floppy disk, you should not have to change the directory, so don't
worry about this. Just press ENTER. If the program has not already been named,
you will be asked to name it. The name must consist of eight or fewer characters
(no punctuation or spaces are allowed). After you type in the name and press
ENTER, ABACUS writes the program on your disk, using the name you typed,
together with the suffix ".ABA" to tag it as an ABACUS program file. You can use
this name to retrieve the program later. Caution: lf you wse the name of a file tbat
is already on the disk, your current program uill be written in place of the one that
bad that name originally, and the latter will be lost.
Quitting
To exit from ABACUS, simply press ESC. To prevent quitting by accident, ABA-
CUS will ask if you really want to quit. Respond by typing 'y' for yes or 'n' for no.
If you have a program in memory, ABACUS also asks whether you want to save it
on the disk, to make sure that you don't accidentally quit without recording your
work. If you're saving to a floppy disk, you can get your disk back by pressing the
button next to the slot into which you put it.
Directory: A:\
Press ENTER to proceed, ESC to abandon, or type a new directory
CrassrcaLPReorcatrLocrc: Ur{orcroasrlrry 279
and pressing ENTER again, you will see a display of the names of the programs
on file. There is a highlighter in this display which can be shifted from name to
name by using the arrow keys on the right side of your keyboard. Move the
highlighter to the name of the program you want to call up and then press ENTER.
The program will then appear on the screen, and you can run it or revise it as you
like. If you revise it, be sure to save the revised version onto the disk before
you quit.
You can also delete programs from the disk by selecting KILL and pressing
ENTER. Notice that KILL is different from ERASE. ERASE removes a program
from the screen but does not affect the disk. KILL deletes a program from the disk
but does not affect what is on the screen.
Experiment and enjoy!
1. This program adds one to the quantity in register A and then halts. Save it
under the name 'PLUS1'.
A+
2. This one repeatedly adds one to the quantity in register A and never halts. Save
it under the name 'COUNTS'.
3. This one empties register A and then halts. Save it under the name 'EMPTY'.
280 CHAPTER 1 0
4. This one empties register A and then just keeps on running without effect. Save
it under the name 'IDLER'.
5. This one adds one to each of registers A and B, then halts; save it under the
name'EACH1'.
0 z
6. \When started with register C empty, this program adds the contents of register
B to register A while retaining the contents of register B. Save it under the name
.ADD'.
I
A+
3 1
∩﹀
4.
Ce
′0
11
Ae
4,
0
12
Ce
│││
10 9
│■ │││││:│■ ■│.・ ■ │││■
・
││‐ ‐
││‐ │・ ││.││.│. ││,│lt‐ │,‐ │‐ ││‐
‐ ‐
│‐ │,.│.(・ │.(│‐
...‐
│‐
│‐
││.1‐
..:‐
│‐
│‐
││.│
1. Write a program that adds two to the number in register A and then halts. Use
only one register. Save this program under the name 'PLUS2'.
2. Often in computer programs, the numbers 1 and 0 are interpreted as the
answers "yes" and "no," respectively.3 \7rite a program which, started with
any natural number in register A, answers the question "Is this number even?"
If the number is even, it signals this by stopping with a 1 (yes) in register A. If
r The correspondence is, of course, arbitrary; we could let 1 be "no" and 0 be "yes,"
or adopr some other convention.
282 (lHAPTER 1 0
the number is odd, it signals this by stopping with a 0 (no) in register A. (Hint:
This program can be done with a single register, but its flow chart will contain
several nodes.) Save this program under the name 'EVEN'.
3. \Write a program that determines for any natural number whether it is evenly
divisible by three. That is, if it is divisible by three, the program should halt
with a 1. in register A; if not, the program should halt with a 0 in register A.
Use only one register. Save the program under the name 'DIV3'.
4. Write a one-register program that answers the question "Is this number less
than four?" That is, when started with a number in register A, it outputs L if
that number is less than four and 0 if it is four or greater. Save this program
under the name 'LESS4'.
5. Write a program that doubles the number in register A. (Hint: Use two regis-
ters; when the program is started, register B must be empty.) Save this program
under the name 'DOUBLER'.
6. Write programs to do truth tables for negation, conjunction, and disjunction.
In a computer, truth values are typically represented by 0's and 1's. Ve'll
follow the usual convention of using 0 for false and 1 for true. A truth table
for conjunction, for example, will work like this:
If registers A and B (representing the two conjuncts) both contain 1 as
input, then the program should halt with 1 in register A (which we
conventionally declare to be the answer register) and nothing in any of
the other registers. (That is, if both conjuncts are true, the table yields
true.) If some other combination of 0's and 1's is in registers A and B,
then the program should halt with 0 in all registers. (Any other combi-
nation of truth and falsity yields falsehood.) Don't worry about what
the program does with numbers greater than 1, since for a truth table
we simply won't allow larger numbers as input.
The negation program should reverse the value in register A. The disjunction
program should halt with a 1 in register A and 0 everywhere else if the input
of either register A or register B was 1; otherwise, it will halt with 0 in all
registers. Try to make these programs as simple as possible, using as few
registers as possible. Save them under the titles'NEG', 'CONJ', and 'DISJ'.
7. Write a program for the operator neither . . . nor. (Hint: This program might
first compute disjunction as in problem 6 and then reverse the truth value of
that computation.) Save this program under the label 'NOR'.
8. Write a two-register program that empties the contents of register A into
register B, then empties the contents of register B into register A, then empties
the contents of register A back into register B again, and so on infinitely. Make
the program run infinitely even if both registers are initially empty. Save it
under the name 'SISYPHUS'.
9. !7rite a program that answers the question "Are these two numbers equal?"1
that is, it compares the contents of registers A and B and outputs a 1 in register
A (with register B empty) if they are equal but 0 in both registers if they are
unequal. Use only two registers. Save this program under the name 'EQUAIJ.
CrnssrcalPRrorcare Locrc: Ut'togctoaetury 283
10. Write a program to exchange the contents of registers A and B. !flhen the
program halts, all the other registers (if any) should be empty. Save this pro-
gram under the name 'EXCHANGE'. (Hint: Use three registers.)
11. One important arithmetical function is min (short for minimum), a two-place
function which, for any numbers x and y, gives as its value the smaller of the
two. If both are equal to some quantity n, then min gives z as its value. More
concisely:
min(x, Y) = x, if x 1Y
min(x, y) = y, if x > !
\7rite a program that computes min. Given any two numbers x and y as input
in registers A and B, your program should eventually halt with min(x, y) in
register A and all other registers empty. (Hint: Use three registers.) Save this
program under the name 'MIN'.
12. \il7rite
a program that computes the max (short for maximum) function, which
is defined as follows:
We have been examining the concept of an abacus in order to clarify the notion of
an algorithm. Specifically, we shall claim that any algorithm can be represented
as an abacus program, and any abacus program is an algorithm. In this sense,
algorithms are just abacus programs, and vice versa. This claim is a version of
Church's thesis.
But why should we accept Church's thesis? The answer is bound up with
some remarkable developments in the history of logic. In the earlier section on
284 CHaprun 10
But the numbering below is not an encoding, since here the number 2 is used
ambiguously to stand for two different states:
lfashington Oregon California
212
But, while the definition of the term "encoding" excludes ambiguities, it admits
the possibility rhat two different numbers from the same code might stand for the
same object, just as a person might have two different names. This is of little
importance, however; the codes we consider will generally assign unique numbers
to objects.
Encodings may be infinite as well as finite. Suppose, for example, that we
wish to encode the following infinite series of formulas:
B -B --P, ---P,...
Perhaps the most natural ways to do so are
R -R --R ---P,...
0123
In each case, there is an algorithm that outputs a code number given a formula,
and a second algorithm that outputs a formula given a code number. To go from
formulas to code numbers (i.e., to encode the formulas), the algorithm for the first
encoding is
If the series of base ten digits that express an even integer n appears in the
decimal expansion of the number fi, then let a encode itself; if not, let it be
encodedbyn-1.
To illustrate, consider the number fourteen. The series of base ten digits that ex-
press this number is'L4', that is, a '1' followed by a '4'. Now the decimal
expansion of n is an infinite nonrepeating decimal, which begins as follows:
3.1,41,59. . . . Sequences of digits appear within ir more or less randomly. The
sequence '14' appears here (the second and third digits), and so in our numbering
14 numbers itself. In general, each even number will be numbered either by itself
or by the odd number before it. But we may not be able to tell which, without
completing an infinite operation. For example, consider the number 217271944.
Does the base ten sequence which expresses it-namelS '2727944'-appear in
the decimal expansion of n? Maybe. If we start cranking out the digits of n, we
may eventually find it. But then again we may not. If in fact the base ten sequence
never turns up, so that the code for 2,747,944 is2,747,943, we'd have to search
through the infinite decimal expansion of n to find out. Thus calculating a code
from a number involves a potentially infinite operation. It is therefore noneffective.
(Note that calculating the number from its code, however, is an effective proce-
dure, assuming that we know that the code is a code. For if the code number is
even, then it stands for itself, and if it is odd, it stands for itself plus one.)
It's not that our sample numbering is ill defined. The rule defining it is per-
fectly clear and exact. But it is not effective and therefore does not count as an
encoding.
CLassrcarPReotcnre Loctc: Ut'lorctoaatrtrv 287
i. If aformula is a sentence letter, its code number is iust the code number
for that letter.
ii.If a formula is a conditional (D -+ Y, its code number is four times the
code number for O plus the code number for Y. (Thus' for example, the
code number for 'P - Q' is (a x 1) + 2 = 6.)
This encoding of formulas does not specifically represent the conditional sign, but
since that sign is the only logical operator available in this fragment of proposi-
tional logic, there is no need; we know in advance that every compound formula
(i.e., formula represented by a number greater than three) is a conditional.
To decode the number for a conditional, we divide by 4; the quotient is the
number of the antecedent, and the remainder is the number of the consequent.
Thus the number 11 represents the formula'Q- R', since 11 divided by 4is2
(the code number for'Q') with a remainder of 3 (the code number for'R')'
Using this encoding, we can now program ABACUS to perform some simple
logical tasks:
1. rirrite a program which, given the code for a conditional in register A and code
nu-beri foi sentence letters in registers B and C, answers the question "Is the
sequent consisting of the premises represented in registers A and B and the
conclusion represented in register C an instance of modus ponens?" That is, it
halts with a I in register A and all other registers empty if this sequent is an
instance of modus ponens and 0's in all registers if not.
2. Generalize the program from problem 1 to test the validity of any two-premise
sequent expressible in this fragment of propositional logic. This involves testing
noi only for modus ponens, but also for hypothetical syllogism, for material
implication, and for conclusions that are tautologous or merely repeat a premise.
288 CHaprun 10
church's thesis tells us that any algorithm can be carried our by an abacus. But
since an abacus operates only on counters, any data we enter into it must first be
numerically encoded so that each potenrial item of input is represenred by a spe-
cific number of counters. Thus, for example, if we wanted an abacus to do tree
tests, we would have to devise a way to encode trees as numbers. This is compli-
cated, but feasible. Individual formulas could be encoded along the lines illustrated
in Exercise 10.4. But an encoding for whole trees would also have to include
symbols representing check marks, X's, and some kind of punctuation to separare
formulas and paths from one another. Given as input the code number of an initial
list of formulas for a tree, the abacus would perform a series of arithmetic opera-
tions on these numbers and, if it halted, output a new number representing the
completed tree.
Such an abacus program would not solve the decision problem for predicate
logic, however; for when given an infinitely prolongable initial list it would run
forever and thus fail to return an answer.
Yet even though for some lists of formulas such a program would be unable
to detect consistency, we mtght be able to do so. The infinite trees we have seen so
far are recognizable as infinite by the patterns of repetition they exhibit. we don't
have to carry them out very far before we notice this. If all infinite trees exhibited
such repetition, and if we could always recognize it, then the tree test coupled with
our repetition-recognizing ability tuould, ar least in principle, enable us to solve
the decision problem. For once we recognized (via the repetition) that a tree was
going infinite, we would know that it must contain at least one open path, and we
could deduce that the initial list was consistent without finishing rhe tree.
But, alas, not all infinite trees are regularly repetitive. some grow without
any recognizable pattern. Their apparent randomness foils the attempt to make
the tree test into a decision procedure by coupling it with a test for pattern
repetition.
Thus we see that our inability to create a decision procedure using the tree
test is not due simply to the fact that some trees are infinite. The deeper source of
the failure is our inability to know for some trees whether they are infinite or not.
More generally, the problem is our inability to recognize whether or nor a given
instance of a potentially infinite process is actually infinite.
so let's try another approach. Given an abacus program that constructs trees,
we might, instead of looking for patterns of repetition in the trees themselves,
CrassrcarPReotcarr Loctc: Ut\orctoeatrtrv 289
simply ask: "Does this program eventually halt, when given as input the code
number of a particular initial list of formulas?" If the answer were yes, then we
could simply run the program to 6nd out whether or not the initial list was
consistent; and if the answer were no, then we would know that the tree was
infinite and hence that the initial list was consistent. Either wan we'd have our
answer.
This approach brings us to the halting problem, the problem of creating an
algorithm that decides for each program and input whether or not that program
halts given that input. A solution to the halting problem, then, would give us a
solution to the decision problem.
However, like the construction of trees, the running of an abacus program is
a potentially infinite process. Some abacus programs halt with a given input after
a finite number of steps; others run forever. Among the nonhalters, some are
readily recognizable as nonhalters by the patterns of repetition they exhibit; others
exhibit complex, nonrepetitive behavior. With the latter, we may find ourselves in
'Sfe
a state of ignorance like that produced by infinite nonrepetitive trees: seem to
have no way of knowing whether the process would end if only we were to carry
it a few (or a few trillion) steps further, or whether it never ends. It seems unlikely,
then, that we could ever produce an algorithm to decide whether or not programs
halt. But can we be sure?
Astonishingly, we can. In this section we show that the very idea of a solution
to the halting problem is self-contradictory. For if we had an algorithm (i.e., abacus
program) that solved the halting problem, we could convert it into what we shall
call a reverse halting program. But, as we shall see, a reverse halting program is a
logically impossible object. Thus we conclude that the halting problem is unsolv-
able, and with that conclusion this section will end. By then we may strongly
suspect that the decision problem is likewise unsolvable-but we will not yet
have proof.
The proof comes in Section 10.5. Here we reverse direction. Having seen
that a solution ro the halting problem would also be a solution to the decision
problem, we consider the converse: Would a solution to the decision problem also
te a solution to rhe halting problem? The answer is yes; thus, in essence, the
decision problem and the halting problem are one and the same.'We prove this by
showing how each abacus program and input can be translated into a sequent of
predicate logic. A test to decide the validity of each such sequent would determine
ior the corresponding program and input whether or not that program halts with
that input-and so would solve the halting problem. Having already seen that
the haliing problem is unsolvable, however, we must then infer that there can be
no algorithmic test to decide the validity of each of these particular program-
representing sequents. And this means that there can be no general solution to the
decision problem.
In outline the central argument of this section and the next is as follows:
ltxyz+Rr*
You are familiar with the character for zero ('0'); the variables 't','x','y', and,z';
the arrow representing the conditional; the predicate letter 'R'; and the successor
symbol "'. The asterisk ('o') is added to make new variables in case we need more
than the four listed here. Appending the asterisk to any variable produces a new
variable. (Thus, for example s'26**'p' is a variable different from the variables 'x',
'xo', and 'rc* "'.) The final mark, the period, is a punctuation mark used to indicate
the end of a program statement. Numbers (to designate register contents or arrows
in the program) are written in this notation as the zero character followed by the
appropriate number of successor marks.
Now to develop a numerical code for programs written in this notation,
we first assign each character in the notation a digit from 0 to 9. Suppose we do it
this way:
R
∩ ︶ ハU
︱ ”
χ う4
Z
ツ
6
.
3
ξJ
Any program can be encoded in this way. Moreover, the processes of encod-
ing and decoding are both clearly algorithmic. Of course, not all numbers repre-
sent programs. The number 1.23, for example, represents the nonsense sequence
'txy'. But that's fine; code numbers need not be consecutive. In discussing the two
kinds of halting problems, we will focus on what happens when we input code
numbers for programs into programs. Keep in mind that the encoding of programs
illustrated here is just one of infinitely many ways of encoding that are possible; it
merely provides a concrete understanding of what an encoding of abacus pro-
grams looks like.
'We
are now ready to give a precise definition of what it means to solve the
halting problem. (NOTE; In tbis definition, and in the ensuing discwssion and
proofs, when we say of an input or output that it is placed in some register or
registers, we assume unless otherwise specified that all other registers are empty.)
The number i represents the input provided to the program numbered e in register
A. \X/e have specified the registers in which the inputs i and e and output of GH
are to be placed merely for definiteness. A machine that in effect solved this prob-
lem but took input or produced output elsewhere could easily be converted into a
machine that fits our definitions, by adding operations to transfer the input or
output to the appropriate registers. Likewise, the specification of the output as
being 0 or 1 is inessential. Any output scheme would do, so long as there was an
effective way to tell which outputs stood for halting and which for nor halting. But
if there were an effective way to do this, then there would also be a way ro program
the machine to convert this nonstandard output into 0's or 1's as described earlier.
(We might, for example, have as output any positive integer, the even ones standing
for halting and the odd ones standing for nonhalting; then we would simply append
a test for evenness to get the ourput to be either 0 or 1.) So, if there is any program
that in effect solves this problem but doesn't fit our definitions, it can easily be
converted into a program that solves the problem and does fit the definitions.
'We
are now poised to prove that a solution to the halting problem is impos-
sible. To do this, we will show that if some abacus program did solve the halting
problem, it would enable us to create a reverse halting program, the very idea of
which is self-contradictory. The concept of a reverse halting program is convo-
luted; so rather than explaining it straight off, we'll first examine a simpler analog:
Bertrand Russell's barber.'
Consider an adult male barber living in a certain village, who for any man
of the village, shaves that man if and only if the man does nor shave himself. This
seems, on the face of it, perfectly possible. Each of the village men either shaves
himself or is shaved by the barber, but not both. But trouble comes when we ask
about the barber himself. Since for any man M of the village, the barber shaves M
if and only if M does not shave himself, and since the barber himself is a man of
the village, it follows that he shaves himself if and only if he doesn't shave himself.
This is a contradiction, which shows that the very idea of such a barber is incoher-
ent; no such barber can exist.T
6 I call this Russell's barber, because it was Russell who invented the example,
though for a different purpose than it serves here.
7 The contradiction is the statement
(" ) The barber shaves himself if and only if he doesn't shave himself.
which has the form 'S * -S'. This form is inconsistent, as you can check with a
truth table or tree. Some people might be bothered by the fact that this contradic-
tion is not in the usual 'P & -P' form. If you are one of these people, consider the
following argument, which deduces a standard form contradiction from state-
ment ("):
Suppose for reductio that the barber shaves himself. Then by (',) and
modus ponens he does not shave himself. Hence he both does and does not
shave himself.
Hence we must reject the reductio hypothesis, and so we have shown that the
barber does not shave himself. But then again by (") and modus ponens he does
shave himself. So it follows from (") alone that the barber both does and does nor
shave himself.
CrassrcarPnrorcarrLocrc:Unorctoagtlttv 293
Now a reverse halting program is like Russell's barber. But whereas the bar-
ber's task is to shave, the task of a reverse halting program is to respond to pro-
gram code numbers given to it as input. Its response is to be the opposite of what
the program indicated by the code number does when given its own number as
input. That is, if the encoded program halts when given its own code number
as input, the reverse halting program runs forever with that same code number as
input. If the encoded program runs forever when given its own code number
as input, the reverse halting program halts when given that number.
Consider, for example, the program P mentioned earlier:
RrOr'* Rr'1x
Rr00 - Rr'20
P subtracts one from the input and halts-or, if the input is zero, halts without
changing anything. Thus it halts no matter what input it is given. So, in particular,
P halts if given its own program number, 61.0275617072961.00561,707709, as
input.
!7hen a reverse halting program is given a program number, however, it does
just the opposite of what the program it stands for does with it. So, since P halts
when given its own number, 61,0275617072961.00561.707709, as input, a reverse
halting program would run forever given this number as input.
Thus, like Russell's barber, who does for each man the opposite with respect
to shaving of what that man does in relation to himself, a reverse halting program
does for each program the opposite with respect to halting of what that program
does in relation to its own program number. More precisely:
Now RH, being by definition an abacus program, must itself have a code number.
And, just as we see that Russell's barber is an impossible object by asking whether
or not he shaves himself, so we can see that RH is an impossible object by asking
what it does when given its own code number as input:
that there is
sorne when given a
only if the
This establishes premise 1 of the argument outlined earlier. Our next task is to
prove premise 2-that is, to show that if there were a solution to the general
halting problem, that would enable us to build a reverse halting program.
METATHEOREM:Ifthcrc is an a
ing problem, then there is a reverse halting progra
PRooF: Suppose for conditional proof that some
solves the halting problem. Then there is an encoding E of abacus
programs such that given any program number e from E and
number I,
△︻ e
Ce GH
│││││● │││
│││││■ ‐■■│
The Program G
‐
‐rし よξ tiSo when‐ startcd with any program number πaSinputin
tei‐
rettster
‐
A, . . .
lr.‐ ‐ ■.´ ‐‐ . ‐
│‐
‐‐
・
│‐ ..‐ ‐ .│‐
answers the question
│.‐
as input?"
GH十 we can also build a
,
GH+ A+
The Program RH
lё in.then thcre is
■
│lr.│││││‐ .■ ‐
now use the two metatheorems just proved to draw the following corol-
'We
lary, which is the major result of this section (this is statement 3 in the outline
given earlier):
In particular, this means that we can't solve the decision problem for predi-
cate logii simply by applying a halting test to a trees program-for the very good
296 CHapren i0
reason that a halting test is an impossible object. But maybe there is some specific
halting test for trees programs, or maybe there is some other sort of algorithm
altogether that could decide the validity of sequents of predicate logic . . .
No such luck.
is decidable.
an
determines for
inputs,
8 Specifically, our proof in this section applies to predicate logic with function sym-
bols. l7hether or not the identity predicate is included makes no difference, since
we make no use of the identity predicate in our proof. In Section 10.7 we modify
the proof to cover predicate logic without function symbols but with the identity
predicate.
Cussrcar PRrorcarr Loctc: UtorctoABlllrY 297
│■ │
│■ ││ ■■│
│■
││■
│ .■ ■
The Prograrn H
│‐
■ ││
■problem.QED
■
298 Cuepren 10
To cinch the proof of the metatheorem on which this corollary depends, we need
to confirm that in fact we can effectively define the sort of associated sequent
described earlier. Here is the definition:
The associated sequent for a program with r registers and input z has two
sorts of premises:
1. If there are one or more exit arrows, construct for each exit arrow a
statement of the form 3t1x, .. . lx, R t e xt .. . x,, where e is the arrow
number, t is a variable representing time, and x, . . . x, are variables indi-
cating the contents of the registers. This statement says that the machine
reaches arrow number e at some time with some input. Then make the
conclusion the disjunction of all such statements. (The conclusion thus
says that the program reaches one of the exit arrows.)
2. If there are no exit arrows, let the conclusion be the negation of the first
premise so that the sequent is invalid.
e We assume they are empty because the halting problem (as defined in Section
10.5) concerns inputs to register A onl6 and we are here designing associated
forms so that a test for their validity would solve the halting problem.
CLassrcarPReorcaruLoctc: Utlorctoastrtlv 299
ハυ
R005 (Machine state when program begins)
Vr(Ril0 - Rr'20)
VlVx(Rl0r'* R/'1x)
- Rr'1x)
Vr(Rr20 R/'20)
YtYx(Rt2x'- (Superfluous)
1t1xPttlx (Halt)
This program, of course, halts with input 5 (it iust subtracts one and stops); ac-
cordinglS the associated sequent is valid. (The conclusion follows from the first
and third premises.) If, however, instead of 'R005', we made the first premise
'R000', then the sequent would be invalid. This, too, is as it should be, for with an
input of 0, the program does not halt; it runs forever, doing nothing.
Notice that for each moment the program is running, the premises of this
sequent imply a true description of the state of the abacus at that moment. Trivi-
ally, for example, these premises imply
R005
which describes the machine state at time 0, simply because this is the first premise.
This conclusion, together with the third premise, implies
R114
which describes the state of the machine at time 1, which is the point at which the
program halts. If the first premise were 'R000' instead of 'R005', then the premises
would imply the following infinite set of true descriptions:
Rsecr...c,
whence it follows validly that
3t1yr.. . fy,R t eyt. . . y,
6f‐ the‐
‐
・
sequent is valid.
Thus P halts given input I if and
i is valid. QED
LEMMA 2: For any time t, if a program has not halted before t, the
premises of its associated sequent validly imply a true description of
time f.
pRoor: The conclusion is a universal quantification over times, which
are a linearly ordered series. Thus we can proceed by mathematical
induction.
Basis Case:
true that (if a program has not halted before f + 1, the premises of its
t+ 1).
for conditional proof that if a program has not halted
before r, the premises of its associated sequent validly imply
the true description of time r. (This is the inductive hypothesis.)
Now since what we are trying to prove is a conditional, we
make another conditional proof hypothesis:
Suppose for conditional proof that a program has not
halted before time / + 1. Then clearly it has not halted
before time /.
ponens)
ply the true description of time /. Now since it
halted at time f, the machine must have been on some
This completes the proof of the undecidability of predicate logic with function
symbols. We have shown that there can be no algorithmic procedure for determin-
ing whether or not any sequent of predicate logic is valid. Some procedures (like
the tree test) work for some sequents, but no procedure can work for all. No
matter what algorithms we create, no matter how sophisticated or fast we make
our computers, there will always be logic problems that we cannot solve.
And we have just proved that-logically!
Vx3ySyx
which means (given a domain of numbers) that every number has at least one
successor and
VtVx(RtOx - Rr'0x')
for example, becomes
VeVyVlVx(S rv
- (Syt
- (Rr0rc * Ry0z)))
This procedure eliminates all occurrences of "' but adds a new universal quantifier
and a new occurrence of 'S' in the antecedent of a conditional for each so that
formulas get cumbersome. That's why we didn't define associated sequents this
way in the first place. Use of the successor function symbol makes both the abacus
programming notation and the definition of an associated inference much simpler.
Even though these new-style associated sequents are more cumbersome, they
work exactly the way the old ones did, and the undecidability proof of Section
10.6 can be redone, by substituting the new sort of associated sequent for the old
sort in the proof. This changes it from a proof of the undecidability of predicate
logic with function symbols to a proof of the undecidability of predicate logic
without function symbols but with the identity predicate.
In fact, predicate logic remains undecidable if we eliminate the identity pred-
icate as well, but we will not show that here.10 In sum, the identity predicate and
function symbols make no difference; predicate logic itself, with or without either,
is undecidable. But why?
The answer, as the infinite trees of Section 7.4 suggest, is that the only counter-
examples for some invalid sequents are counterexamples with infinite domains-
domains constructible only by an infinite process. The ability to describe such
infinities is one of predicate logic's virtues; without this ability, predicate logic
would be of little use, for example, in higher mathematics, where it has in fact
found many applications. But with this virtue comes the corresponding vice of
undecidability.
For some uses, however, we do not need the full power of predicate logic: A
mere part or fragment of predicate logic will do. We have already seen that one
such fragment, propositional logic, is decidable. Are there other decidable
fragments-fragments that include more than iust propositional logic?
Indeed there are. One is the fragment obtained by dropping the existential
quantifier from predicate logic without function symbols and requiring all univer-
sal quantifiers to take wide scope (i.e., not to occur within the scopes of operators
that are not universal quantifiers). The resulting fragment of predicate logic,
10 One way to do it is to use the sort of associated sequents just illustrated but
eliminate the identity predicate by defining it in terms of the other predicates used
in those sequents-namely,'R'and 'S'. The technique for eliminating the identity
predicate is explained in W. V. Quine, Philosophy of Logic (Englewood Cliffs,
N.J.: Prentice-H a11,1.970), pp. 63-54.
304 CHAPTER 1 0
though quantified, is, like propositional logic, decidable. It never generates infinite
trees, because once each formula's initial string of quantifiers is instantiated for all
names akeady on the path, no new names can appear and the tree can be finished
simply by repeated application of the propositional rules. This is true even when
we negate a universal conclusion to test for validity, which will yield an existential
formula by -V. Once all the initial quantifiers of this formula are instantiated-
by repeated steps of -V and f 1r-this formula can generate no new names. Re-
stricting predicate logic in this way would, of course, limit the things we c,ould say,
but it might be adequate for some uses.
Other fragments of predicate logic are decidable as well so that if we restrict
our attention to certain classes of problems, we may sometimes be sure of getting
an answer even if we are using quantifiers. Much technical work has been devoted
to sorting out decidable fragments of predicate logic from the undecidable ones,
but we shall pursue this theme no further.
Before we leave the topic of undecidability, one small but significant doubt
needs airing. The proof of premise 4 in our outline of the undecidability proof
assumed Church's thesis. But Church's thesis is not a logical truth. It is a well-
confirmed empirical hypothesis-so well confirmed that by now it in effect defines
the notion of an algorithm. It is conceivable, however, that we might some day
discover something that is recognizably a computational process but which nev-
ertheless is not representable as an abacus program. What we would say then
would depend on whether we think of Church's thesis as an empirical hypothesis
or a definition. If we think of it as an empirical hypothesis, we would say that it
proved to be false. If we think of it as a definition, then we would simply announce
the discovery of nonalgorithmic computational processes.
However we described it, such a discovery would undermine our proof of
the undecidability of predicate logic, but it would by no means show that predicate
logic is decidable. The new computational procedures might be no more capable
of solving the decision problem than abacus programs are. This whole scenario,
however, is highly speculative.
Exercise 10.7
Is the fragment of predicate logic that results from allowing only existential quan-
tification and requiring these quantifiers to take wide scope decidable? \7hy or
why not?
11 Actually, having banned existential quantification, we would not really have the
I rule for trees; we would need instead a modified -V rule which combined the
functions of our current -V and I rules.
lll「
P′ ART」││「
LE:BNIZ:AN MODALLOGiC
C月 ハ Pア fR t,
307
308 Cuapren I l
example, these sentences, all of which involve the application of modal operators
(in the broad sense) to the single proposition 'People communicate':
Alethic Operators
Deontic Operators
There are, of course, many more operators in each category. And some of those
listed, such as 'it is possible that' and 'it could be the case that' are, at least in some
contexts, semantically identical or synonymous. \ff/ith the exception of the opera-
tors expressing propositional attitudes, all of those listed here are monadic; they
function syntactically just like the negation operator 'it is not the case that', prefix-
ing a sentence to produce a new sentence. Thus, for example, the operators 'it is
necessary that', usually symbolized by the box 'tr' and 'it is possible that', usually
Lrrsr.rrzraN Mooar Locrc 309
symbolized by the diamond sign 'O',1 are introduced by adding this clause to the
formation rules:
If O is a formula, then so are !@ and O<D.
tr@ * -O-@
O@ * _tr_O
That is, it is necessary that O if and only if it is not possible that not-(D, and it is
possible that @ if and only if it is not necessary that not-(D.
There are other duals among these operators as well. Consider the deontic
operator 'it is obligatory that', which we shall symbolize as 'O', and the operator
'it is permissible that', which we shall write as 'P'. These are similarly related:
OO * -P-<D
P@
- -O-<D
That'O' and 'P' should thus mimic 'tr' and '0' is understandable, since obligation
is a kind of moral necessity and permission a kind of moral possibility.
There are also epistemic (knowledge-related) duals. The operator 'knows
that' is dual with the operator 'it is epistemically possible, for . . . that'-the for-
mer representing epistemic necessity (knowledge) and the latter epistemic possibil-
ity. (Something is epistemically possible for a person if so far as that person knows
it might be the case.) Symbolizing 'knows that' by 'K' and 'it is epistemically
possible for . . . that'by 'E', we have:
pK@ * -pE-O
pEO .- -pK-O
In English: p knows that <D if and only if it is not epistemically possible for p that
not-<D; and it is epistemically possible for p that <D if and only if p does not know
that not-@ ('p', of course, stands for a person).
There are temporal duals as well. Let 'P' mean 'it was (at some time) the case
that' and 'H'mean'it has always been the case that'. Then:
H@ * -P-O
P@ * -H-O
Here 'H' represents a kind of past tense temporal necessity and 'P' a kind of past
tense temporal possibility. A similar relationship holds between'it always will be
the case that' and 'it sometimes will be the case that' and between other pairs of
temporal operators.
These systematic logical relationships bear a striking resemblance to two
familiar laws of predicate logic:
Vx(D * -lr-@
lx(D * -Vx-@
Are these pairs of dual operators somehow analogous to quantifiers?
Leibniz, who was among the first to investigate the logic of alethic operators, in
effect suggested that they are. His semantics for modal logic was founded upon a
simple but metaphysically audacious idea: Our universe is only one of a myriad
possible universes, or possible worlds. Each of these possible worlds comprises a
complete historn from the beginning (if there is a beginning) to the end (if there is
an end) of time.
Such immodest entities may rouse skepticism, yet we are all familiar with
something of the kind. I wake up on a Saturday; several salient possibilities lie
before me. I could work on this book, or weed my garden, or take the kids to the
park. \Thether or not I do any of these things, my ability to recognize and entertain
such possibilities is a prominent feature of my life. For ordinary purposes, my
awareness of possibilities is confined to my doings and their immediate effects on
the people and things around me. Yet my choices affect the world. If I spend the
day gardening, the world that results is a different world than if I had chosen
otherwise. Leibnizian metaphysics, then, can be seen as a widening of our vision
of possibility from the part to the whole, from mere possible situations to entire
possible worlds.
Possible worlds figure most notoriously in Leibniz's theodicy. God, in con-
templating the Creation, surveyed all possible worlds, says Leibniz, and chose to
actualize only the best-ours. Since ours is the best of all possible worlds, the
degree of evil or suffering it contains is unavoidable-as we would see if only we
had God's wisdom.2
What interests the logician, however, is not how Leibniz used possible worlds
to rationalize actual miseries, but how he used them to adumbrate an alethic
modal semantics. On Leibniz's view:
!@ is true if and only if (D is true in all possible worlds.
'z This has given rise to the quip that the optimist is one who, like Leibniz, thinks
that ours is the best of all possible worlds, whereas the pessimist is one who is sure
of it.
LrraNrrzraNl [\4ooar Locrc 311
and
3 Cf. Martin Heidegger's contention that Dasein (human existence) is its possibili-
ties and thus is more than it factually is; Being and Time, trans. John Macquarrie
and Edward Robinson (New York: Harper & Row), pp. 58, 183-84, 185.
312 CHAPTER ll
icates and names within that domain. In modal logic, we posit many possible
worlds. A model for modal logic, then, should contain many "worlds," each with
its own domain. And because the facts differ from world to world, that model
should assign to each predicate not just a single extension, but an extension in
each world. To keep things manageably (but preposterously) simple, consider a
model representing just three possible worlds, wt, wz, and w3. And (still oversim-
plifying) let's suppose that w, contains exactly four objects, o, F, 1, and 6; w,
contains exactly two objects, B and 1; and w. contains exactly three objects a, 6,
and e:
World Domain
w1 {o, B, "Y,6}
w2 {9,1}
w3 {cr, 6, e)
Now suppose we want to interpret the one-place predicate 'B', which for the sake
of definiteness we may suppose means "is blue." Since a thing may be blue in one
world but not in another, we cannot assign this predicate a single set (the set of
blue things), as we would have in predicate logic. Rather, we need to assign it a
separate set in-or "at" (either preposition may be used)-each world. For each
world u, the set assigned to 'B' at w then represents the things that are blue in ra.
Suppose we assign to 'B' the set {ct, $l in wr, { ) in wr, and {cr, E, e} in w,Then,
according to our model there are two blue things in w, and none in wr, and in u,
everything is blue.
Because extensions differ from world to world (i.e., are world-relative) in
modal logic, a valuation V now must take into account not only predicates, but
also worlds, in assigning extensions. Thus we write
Y('B',wr) = {ct, B}
Y('B'rwr)={ }
Y('B',w.) = {o,6, e}
to indicate that at world w, the set of things that satisfies the predicate 'B' (i.e., the
set of blue things) is {a, B}, and so on.
Truth, too, is now world-relative. BIue things exist in wrbut not in rur; thus
the formula 'lxBx' ought to be true at w, but not at wz. That is, Y ('lxBx' , wr) =
I but Y('lxBx', wr) =F. AccordinglS when we assign truth values to sentence
letters, we shall have to assign each letter a truth value for each world. Let'M', for
example, mean "there is motion." We might let 'M' be true in wrbut not in w2 or
rar. Thus V('M', w r) = T, but V('M', w z) = T ('M', w s) = F.
'We
shall assume, however, that names do not change denotation from world
to world. Thus we shall assign to each name a single object, which may inhabit
the domains of several possible worlds, and this assignment will not be world-
relative. This models the metaphysical idea that people and things are "spread
out" through possibilities, just as they are "spread out" through time. Ifith respect
to time, for example, the name 'John Nolt' refers to me now, but also to me when
I was a child and to the old man whom (I hope) I will become. I occupy many
Lrrsr'irzrax N/oonr Locrc 313
Given any valuation, the following valuation rules describe how truth and falsity
are assigned to complex formulas:
Since the valuation rules are a lot to swallow in one bite, we'll take the proposi-
tional fragment of the semantics by itself first and come back to the full modal
predicate logic later. This simplifies the definition of a valuation considerably:
Here worlds are like the (horizontal) lines on a truth table, in that each is distin-
guished by a truth-value assignment to atomic formulas-though not all lines of a
truth table need be represented in a single model.
Consider, for example, the following valuation of the formula '(V v !7)'
which we may suppose means "Sam is virtuous or Sam is wicked":
,W, = {1,2,3,4}
"f('V', 1)= T Y(.!(/', 11= P
Y('V', 2) = P Y('W',2) =fi
Y('V', 3) = f' 7('W', 3) : T
Y('V',4) = f' 7('W', 4) =T
The "worlds" here are the numbers L1 21 3, and 4. (In a model, it doesn't matter
what sorts of objects do the modeling.) In world 1, 'V' is rrue and '!7' is false-
that is, Sam is virtuous, not wicked. In world 2, Sam is neither virtuous nor wicked.
And in worlds 3 and 4, Sam is wicked, not virtuous.T Our model represents the
situation in which Sam is both virtuous and wicked as impossible, since this situ-
ation occurs in none of the four possible worlds. In other words, only three of the
four lines of the truth table for 'V v'W' are regarded as possible. This is arguably
appropriate, given the meanings we have attached to 'V' and ''W'.
7 In a sense, world 4 is redundant, since from the point of view of our model it
differs in no way from world 3. But this sort of redundancy is both permissible
and realistic. It may, for example, represent the idea that world 4 differs from
world 3 in ways not relevant to the problem at hand; for example, Sam may be a
sailor in world 3 but not in world 4. Of course, if the model were truly realistic, it
would contain many more worlds representing many such irrelevant differences,
but we are simplifying.
LrrsNlrzraNr Mooar Locrc 317
To understand more about how this model works, we must consider the
valuation rules for propositional modal logic (rules 4-8 and 11-1'2 above). Ac-
cording to rule 6, for example, the statement 'V v'W' has the value T in a world ra
if and only if either 'V' or '\(/' has the value T in that world, and it is false
otherwise. Thus this statement is true in worlds 1, 3, and 4, but false in world 2.
The rules for the other truth-functional propositional operators ('-', '&', '-', and
'*') are all similarly relativized to worlds.
The real novelty, though, and the heart of Leibniz's insight, lies in rules 11
and 1.2. Consider, for example, the statement 'tr-(V 6a Ii7)', which according to
our interpretation means "it is necessarily the case that Sam is not both virtuous
and wicked." According to rule 11, this formula is true at a given world w if and
only if the statement'-(V & W)' is true in all worlds. Now in our model '-(V &
'S7)'
is in fact true in all worlds. For there is no world in which both 'V' and 'W'
are true; hence by rule 5, 'V & !(/' is not true in any world, and so by rule 4,'-(V
& \7)' is true in each world. This means by rule 11 that 'n-(V & W)' is true in
every world.
SimilarlS the statement'OV'("it is possible that Sam is virtuous") is true
'Whichever
in all worlds. For consider any given world aa. world z is, there is
some world u (namely, world 1) in which'V'is true. Hence by rule 12,'OV'is
true in w.
Notice also that since 'O V' is true in all worlds, it follows by another appli-
cation of rule 11 that'trOV' ("it is necessarily possible that Sam is virtuous") is
true in all worlds. In fact, repeated application of rule 11 establishes that'!trOV',
'!nEOV', and so on are all true at all worlds in this model. The following meta-
theorem exemplifies the formal use of modal semantics; use it as a model for
Exercise 11.2.L:
Exercise 11.2.1
Consider the following model
(171年
={1,2,3)
γ (`P',1)=T γ(`Q',1)=F γ(`R',1)=T
Ⅵ `P',2)=F γ(`Q',2)=F γ(`R',2)=T
γ (`P',3)=T γ(R',3)=T γ(`R',3)=T
318 CHaprrn 1 1
Using the valuation rules, prove the following with respect to this model:
DEFINITION Two formulas are equivalent iff they have the same truth
value at every world on every valuation of both.
DEFINITION A counterexample to a sequent is a valuation containing a
world at which its premises are true and its conclusion is false.
e Here we contradict Leibniz, who thought that actuality raas something absolute-
namely, whatever it was that God added to our possible world in order to create it
(ours was, according to Leibniz, the only world God created). For a fuller discus-
sion of the indexicality of actuality, see David Lewis, Oz the Plurality of 'l(/orlds
(Oxford: Basil Blackwell,'l986l,sec. 1.9, pp. 92-96.
320 CHeprrn 1 1
The converse, of course, does not hold. What is need not be necessary. The Earth
is populated; but this is nor necessarily the case. (It might cease to be the case
through any of a variety of catastrophic events, and indeed it might never have
happened at all.) To vivify the next metatheorem, think of 'P' as meaning "the
Earth is populated," and think of world 1 as the actual world and world 2 as a
world in which the Earth is barren.
‐
‐
The sequent'・
P卜 │□ P'lis invalidi‐ ・ │.
│■ │
PROOF: Consider the valuatiOn° /whOse set■ ち,OfW611ldζ is
││. │■ .・ ││
is truth in all worlds, and if something is true in all worlds, then it is true in all
worlds that it is true in all worlds. The following metatheorem gives the details:
QED
Worldvariables ('w','Lt','x',afid'y',forexample,inthepreviousmetatheo-
rem) are a pervasive feature of modal metalogic. Each such variable should be
introduced with a metalinguistic quantifier to indicate whether it stands for all
worlds or just some. Variables standing for a particular world may be repeated
later in the proof if there is need to refer to that world again. Early in the previous
metatheorem, for exampl er'w'is introduced (via existential quantification: "there
is a valuation V containing a world w"l to stand for a particular world; later it is
employed several times to refer to that same world. To avoid ambiguity, it is best
to choose a typographically new variable for each quantification. Thus, for exam-
ple, in the same proof, 'y' is used to make a universally quantified statement, and
'u' and 'r' to make separate existentially quantified statements.
Our next metatheorem proves one of the two biconditionals expressing the
idea that'!'and'O'are duals. (The other is left as an exercise below.) In some
systems one of these two operators is taken as primitive and the other is defined in
terms of it using one of these biconditionals.
322 Cnaprrn 1 1
■│ │■ │‐.‐
││││ ││
‐ 9崎
‐
‐
│■ │‐││‐ ´at
that y(□ ∼
Φ ).=
"″
th1l b。 7(‐ ◇
■‐ 0,
″ in
γ Z)■ │■
‐ (∼ ,,‐
・‐
assumcd that
is not
One of the most important consequences of the doctrine that names are rigid
designators is the thesis expressed in the next metatheorem: the neccessity of iden-
tity. Kripke, who popularized this thesis in its contemporary form,10 illustrates it
with the following example. 'Phosphorus' is a Latin name for the morning star;
'Hesperus' is the corresponding name for the evening star. But the morning star
and the evening star are in fact the same object, the planet we now call Venus.
Hence the statement
Hesperus = Phosphorus
is true. Now if names are rigid designators, then since this statement is true, the
object designated by the name 'Hesperus' in rhe actual world is the very same
object designated by'Hesperus'in any other world, and the object designated by
the name 'Phosphorus' in the actual world is the same as the object designated by
that name in any other world. Thus in every world both names designate the same
object they designate in the actual world: the planet Venus. So 'Hesperus = Phos-
phorus' is not only true in the actual world but necessarily true.
Yet this conclusion is disturbing. So far as the ancients knew, Hesperus and
Phosphorus could have been separate bodies; it would seem, then, that it is not
necessary that Hesperus = Phosphorus.
But this reasoning is fallacious. The sense in which it was possible that Hes-
perus was not Phosphorus is the epistemic sense; it was possible so far as the
The sequent says that given that something is E, it follows that something
(that very same thing, if nothing else) is possibly F. This is a consequence of the
fact that the actual world, which we may think of as w- and also w-intheproof,
is also a possible world, so that whatever actually has a property also possibly has
324 Crapren 1 1
it. In the proof, the object which actually has the property F is object ;. Since r has
F in w, I has F in some possible world, i.e., possibly has F. It follows, then, that
something possibly has F. This enables us ro conrradict the reductio hypothesis.
Our final merarheorem shows that from the fact that it is possible some-
thing is F, it does not follow that the world conrains anything which itself is
possibly F. Suppose, for example, rhar we admit that it is (alethically) possible
that there are such things as fairies. (That is, there is a possible world containing
fairies.) From that it does not follow that there is in the actual world anything
which itself is possibly a fairy. The counterexample presented in the following
metatheorem is a formal counterpart of this idea. Think of world 1 as the ac-
tual world, which (we assume) conrains no fairies and world 2 as a world in
which fairies exist. (The fairies are objects 6 and e.) Read the predicate'F'as "is
a fairy."
WOrld
‐
■ Domain
‐
. .‐
‐‐‐
l■ │ 1 {α ,β ,χ )
2 .‐ {α ちβ:χ ,δ ;‐
ё
}
│‐ │││_ ‐
‐││││
‐
wh`:ё
NowT('03rFr',
SO that by rule 1,
2)
1).=■
平
‐
│
γ(`ヨ 受◇Fχ ',1)≠ ^nd
■ br f19m
l which
%ln
^Z
world 1
QED
Notice that in the proof of this rheorem we avoided the question of predication
for nonexisting objects (which we have left unsettled). In this case it is the ques-
tion whether the objects 6 and e, which are fairies in world 2, are also fairies in
LrreNrzraNr lVooal Locrc 325
world 1, where they do not exist. Our valuation rules do not answer this question,
but the sequent 'OlxFx l- lr0Fx' is invalid regardless of how it is answered.
Exercise 11.2.2
Prove the following metatheorems using Leibnizian semantics for modal predicate
logic:
1. The sequent'P F OP' is valid.
2. The sequent'OP l- P' is invalid.
3. The sequent'O(P Sc Q) f OP & OQ' is valid.
4. The sequent'OP & OQ l- O(P & Q)' is invalid.
5. Every sequent of the form O F nOO is valid.
5. Every sequent of the form O!@ F n<D is valid.
7. For any formula O, if @ is a valid formula, then so is tr(D.
8. Every formula of the form n<D * -O-@ is valid.
9. Every sequent of the form trO l- OO is valid.
10. Every sequent of the form !(@ - Y) F (!(D - nY) is valid.
11. Every sequent of the form tr(O - Y) F -O(@ & -Y) is valid.
12. Every sequent of the form -O(O & -Y) l- D(O * Y) is valid.
13. The sequent'nP, P-Q 1- lQ' is invalid.
14. Every formula of the form Eo = ct is valid.
15. Every sequent of the form -o = B l- tr-ct = B is valid.
16. Every sequent of the form Oct = I l- o = B is valid.
t7. The sequent'f36trFx F lVxFx'is invalid.
18. Every sequent of the form VPno l- VPo is valid.
Our model theory (semantics) deepens our understanding of the alethic modal
operators, though to get interesting results we have had to make a metaphysical
assumption or two along the way. Still we have not learned much about possibility
per se. The models we have so far considered are all wildly unrealistic-because
they contain too few worlds; because these "worlds" are not really worlds at all,
but numbers; because their domains are too small; and because we never really
said what the objects in the domains were. In this section we seek a more realistic
understanding of possibility by correcting these oversimplifications.
In Section 7 .2 we noted that, although most of the models we encounter even
in predicate logic are likewise unrealistic (being composed of numbers with artifi-
cially constructed properties and relations) we can, by giving appropriate mean-
ings to predicates and names, produce a natural model. A natural model is a model
whose domain consists of the very objects we mean to talk about and whose
predicates and names denote exactly the obiects of which they are true on their
intended meanings. A natural model for geometry, for example, might have a
325 Craprrn 1 1
domain of points, lines, and planes. A natural model for subatomic physics might
have a domain of particles and fields.r I
A natural model for modal discourse will consist of a set of possible worlds-
genuine worlds, not numbers-each with its own domain of possible objects. And
that set of worlds will be infinite, since rhere is no end to possibilities.
But what ls a possible world?
Leibniz thought of possible worlds as universes, more or less like our own.
But how much like our own? can a universe contain just one object? There is no
obvious reason why not. can it contain infinitely many? It seems so; in fact, for
the century or two preceding Einstein, many astronomers thought that the actual
universe really did. Y/e have akeady said that there is a possible world in which I
am a farmer. Is there one in which I am a tree?
This is a question concerning my essence, that set of properties which a thing
must have in order to be me. what belongs to my essence? Being a professor is
pretty clearly zol essential to me. what about being (biologically) human? There
are fairy tales in which people are turned into trees and survive. Do these tales
express genuine possibilities? Such questions have no easy answers. Perhaps they
have no answers at all.
Philosophers who think that the nature of things determines the answers are
realists about essence. Realists believe that essences independent of human thought
and language exist "out there" awaiting discovery. (rVhether or not we can dis-
cover them is another matter.) opposed to the realists are nominalists, who think
that essences-if talk about such things is even intelligible-are not discovered,
but created by linguistic practices. lXrhere linguistic praitices draw no sharp lines,
there are no sharp lines; so if we say increasingly ourrageous things about me (I
am a farmer, I am a woman, I am a horse, I am a tree, I am a prime number, . .),
there may be no definite point at which our talk no longer expresses possibilities.
For nominalists, then, it is not to be expected that all questions about possibility
have definite answers. (Extreme nominalists deny that talk about possibility is even
intelligible.) The realist-nominalist debate has been going on since the Middle
Ages; and, though lately the nominalists have seemed to have the edge, the issue is
not likely to be settled soon.
To avoid an impasse at this point, we shall invoke a distinction that enables
us to sidestep the problem of essence. \il/hether or not it is metaphysically possible
(i.e., possible with respect to considerations of essence) for me to be a tree,itdoes
seem logically possible (i.e., possible in the sense that the idea itself-in this case
the idea of my being a tree-embodies no conrradiction). contradiction is perhaps
a clearer notion than essence; so let us at least begin by thinking of our natural
model as modeling logical, not metaphysical, possibility.
In confining ourselves to logical possibility, we attempt to think of objects as
-il7hat
essenceless. sorts of worlds are possible now? It would seem thar a possible
11 These would be models for theories expressed in predicate logic, not necessarily
in modal logic.
Lrrsr,JrzraNr [\4ooar Loctc 327
world could consist of any set of obiects possessing any combination of properties
and relations whatsoever.
But new issues arise. Some properties or relations are mutually contradictory.
It is a kind of contradiction, for example, to think of a thing as both red and
colorless. SimilarlS it seems to be a contradiction to think of one thing as being
larger than a second while the second is also larger than the first. But these con-
tradictions are dependent upon the meanings of certain predicates: 'is red' and
'is colorless' in the first example; 'is larger than' in the second. They do not
count as contradictions in predicate logic, which ignores these meanings (see Sec-
tion 9.4).
If we count them as genuine contradictions, then we must deny, for example,
that there are logically possible worlds containing obiects that are both red and
colorless. If we refuse to count them as genuine contradictions, then we must
condone such worlds. In the former case, our notion of logical possibility will be
the informal concept introduced in Chapter 1. In the latter, we shall say that we
are concerned with purely formal logical possibility.
Only if we accept the purely formal notion of logical possibility will we
count as a logically possible world any set of objects with any assignment what-
soever of extensions to predicates. If we accept the informal notion, we shall
be more judicious-rejecting valuations which assign informally contradictory
properties or relations to things. lWe shall still face tough questions, however,
abort *hat counts as contradictory. Can a thing be both a tree and identical
to me? That is, are the predicates 'is a tree' and 'is identical to John Nolt' con-
tradictory? The problem of essence, in a new guise, looms once again. Only by
insisting upon the purely formal notion of logical possibility can we evade it
altogether.
In the next chapter the lovely simplicity of Leibnizian semantics will be shat-
tered, so we might as well allow ourselves a brief moment of logical purity now.
Let's adopt, then, at least for the remainder of this section, the formal notion of
logical possibility.
Now, take any set of sentences you like and formalize them in modal predi-
cate logic. The natural model for these sentences is an infinite array of worlds. Any
set *hatsoeuer of actual and/or merely possible objects is a domain for some world
in this array. The predicates of the formalization are assigned extensions in each
such set in all possible combinations (so that each domain is the domain of many
worlds). Among these domains is one consisting of all the objects that actually
exist and nothing more. And among the various assignments of extensions to
predicates in this domain is one which assigns to them the extensions they actually
do have. This assignment on this domain corresponds to the actual world. (Other
assignments over the same domain correspond to worlds consisting of the same
objects as the actual world does, but differing in the properties those obiects have
or the ways they are interrelated.) If our discourse contains any names, on the
intended interpretation these names name whatever obiects they name in the ac-
tual world; but they track their objects (i.e., continue to name them) through all
the possibilities in which they occur.
328 CHAPTER ll
Leibnizian propositional logic retains all the inference rules of classical proposi-
tional logic but adds new rules to handle the modal operators. Though we shall
examine inferences involving identitS we shall not deal with quantifiers in this
section, since the quantifier rules depend on how we resolve the question of pred-
ication for nonexisting objects. One reasonable way of resolving this question is
to adopt a free logic-that is, a logic free of the presupposition that every name
always names some existing thing. 'We shall consider free logics in Section 15.1,
and we defer the treatment of modal inferences involving quantification to that
section.
The nonquantificational Leibnizian logic that we will explore in this section
adds to the rules of classical propositional logic and the classical rules for identity
seven new inference rules (the names of most are traditional and of various
origins):
Duality (DUAL) From either of OO and -n-(D, infer the other; from eirher
of n(D and -O-O, infer the other.
K rule (K) From n(<D
- Y), infer (D@ * nY).
T rule (T) From D@, infer @.
It is not difficult to show that every instance of each of these rules is valid on a
Leibnizian semantics-and indeed we did this for some of them in Section 11.2
(the rest were left as exercises).
The necessitation rule differs from the others in that it uses no premises but
refers, rather, to theorems established by previous proofs. A theorem is a valid
formula, a formula true in all worlds on all valuations. Therefore, if Q is a theo-
rem, trO and any formula of the form EO may be asserted anywhere in a proof
without further assumptions. When we use the rule of necessitation, we annotate
it by writing its abbreviation 'N' to the right of the introduced formula, followed
by the previously proved theorem or axiom schema employed.
These seven inference rules, together with the rules of classical propositional
logic and the identity rules =I and =E, constitute a system of inference that is sound
and complete with respect to a Leibnizian semantics for the modal propositional
logic with identity-but to show this is beyond our scope. The purely proposi-
LrrsNrrzraN lrz'looal Loctc 329
tional rules (i.e., the ones other than =I, =E, and n=) comprise a logic known as
55.12 This section is largely an exploration of the valid inferential patterns of 55.
'We
begin by proving the sequent'P F O P':
1.P A
2.I n-P H (for -l)
3.1 -P 2T
4. I P&-P 1,3&t
5. -tr-P 2-4 -l
6. OP sDUAL
The strategy is an indirect proof. Recognizing initially that'O P'is interchangeable
with '-tr-P', we hypothesize'a-P'for reductio. Using the T rule, the contradiction
is obtained almost immediately. This yields '-a-P', which is converted into'OP'
by DUAL at line 6.
The rules N and K are often used together to obtain modalized versions of
various theorems and rules. The sequent'n(P & Q) F trP', for example, which is a
modalized version of &E, is proved by using N and then K:
1. !(P&Q) A
2. a((P & Q)* P) N ((P & Q)- P)"
3. tr(P 6c Q).* nP 2K
4. oP 1,3-E
A similar but more sophisticated strategy utilizing N and K yields sequents
involving possibility. Our next example is a proof of ' O P F O (P t, Q)', a modalized
version of vI. Here we apply N to the theorem '-(P v Q) .- -P', the contrapositive
of 'P - (P v Q)', which in effect expresses vL (This strategy of applying N to
contraposed nonmodal versions of the modal sequent we want to prove is typical
when the modality involved is possibility.)
1.OP A
2. !(-(P v Q)* -P) N (-(P v Q)- -P)"
3. a-(P v Q)- n-P 2K
4, -a-P l DUAL
5. -r-(P v Q) 3,4 MT
6. O(PvQ) 5DUAL
Note the use of the derived rule modus tollens at line 5. Derived rules for classical
propositional logic (see Section 4.4) are all available in Leibnizian modal logic.
N and K are used together once again in this derivation of the theorem
'F O-P - -trP':
12 The name originates with the logician C. I. Lewis, whose pioneering work on
modal logic dates from the first few decades of the twentieth century. Lewis ex-
plored a number of modal systems, which he christened with such unmemorable
labels. InexplicablS the labels stuck.
1r This theorem is problem 2 of Exercise 4.4.2.
ra See problem 6 of Exercise4.4.2.
330 (二 HAPTER l l
◇∼P H(br→ I)
2
∼□∼∼P l DUAL
3
□P― →□∼∼P 3K
5
∼□P 2,414T
6
◇∼P―→∼□P l-5→ I
Howevet a very different strategy lnay bc used tO prOve the relatcd theorenn
`ト ロ∼P―→∼◇P':
1
□∼P H(br→ I)
2
◇P H(br∼ I)
3
∼□∼P 2 DUAL
4
∼◇P 2-4∼ I
6
□∼P―→∼◇P l-5-→ I
Here, after hypothesizing the theorem's antecedent for conditional proof, we em-
ploy an indirect proof, hypothesizing ' O P' for reductio at line 2. The use of DUAL
at line 3 immediately provides a contradiction, which is recorded at line 4, and the
conclusion follows by easy steps of -I and
-I at lines 5 and 6.
The following proof of the sequent 'tr(P
- Q) l- tr-Q * a-P', which is a kind
of modalized version of modus tollens, displays further uses of N and K:
1。 □(P→ Q) A
2.□ ((P→ Q)→ (∼ Q→ ∼P)) N((P→ Q)→ (∼ Q→ ∼P))
3。 □(P→ Q)→ □(∼ Q→ ∼P) 2K
4。 □(∼ Q→ ∼P) l,3-→ E
5。 □∼Q→ □∼P 4K
The necessitation rule N is used at line 2 with the theorem 'l- (P * Q) (-Q
-P)', which was proved in Section 4.4.|n the use of K at line 3, @ is 'P *- Q' and
-
Y is '-Q - -P', but at line 5 @ is '-Q' and Y is '-P'.
The B rule is used in the following proof of 'O !P F P':
1. 2 . 3 . 4 . 5 . 6
◇□P A
∼P H(fOr∼ I)
□◇ ∼P 2B
□(◇ ∼P―→ ∼□P) N(◇ ∼P→ ∼□P)
□く〉∼P―→ □∼□P 4K
∼□∼□P l DUAL
。
7. 8
∼□◇ ∼P 5,6 MT
□◇ ∼P8c∼ □◇∼P
。
3,7 8ζ I
9. 0
∼∼P 2-8∼ I
P 9∼ E
.
Note the use of N with the previously proved modal theorem'O-P - -trP'at
line 4.
'We
next prove the theorem 'F O OP
- OP', using the 54 rule:
1
◇◇P H(br→ I)
2
□∼P H(f()r∼ I)
3
□□∼P 2S4
4
□□∼P―,□ ∼◇P 4K
6
□∼◇P 3,5-→ E
7
∼□∼◇P l DUAL
8
∼□∼P 2-8∼ I
0
◇P 9 DUAL
11。 ◇◇P― → ◇P l-10-→ I
This theorem can easily be strengthened to the biconditional 'OP * O OP', using
the previously proved sequent'P l- OP' as a derived rule. This biconditional shows
that repetition of possibility operators is in effect redundant in Leibnizian logic.
The same can be shown for necessity operators-that is, '[ lP * ncP', but the
proof is left as an exercise.
As in propositional and predicate logic, we may use derived rules. We will
not, however, bother to name them, since few have widely used names. Instead,
we simply list the previously proved sequent to the right, together with the line
numbers of the premises (if any) that are instances of the previously proved se-
quent's premises. (Rules derived from theorems have no premises, and we cite no
lines for them.) This proof of 'l(P
- Q) l- OP - OQ' uses the previously proved
sequent 'tr(P .- Q) F tr-Q -* D-P' as a derived rule at line 4:
Q◇︱
↓
P
1
A
2
H(fOr→ I)
3
□∼Q H(fOr∼ I)
4
□∼P 3,4-→ E
6
∼□∼P 2 DUAL
︱
7
│ ∼□∼Q 3-7∼ I
9
I ◇Q 8 DUAL
0
◇P→ ◇Q 2-9-→ I
Notice the use of indirect proof with the duality rule to obtain 'OQ'.
As I pointed out in Section 4.4, proof of a sequent establishes the validity of
any formula that shares that sequent's form. Thus, when we use a sequent as a
derived rule, we may use any instance of it. The following proof of the sequent
'-a=b F tr-a=b'utilizesthe previouslyproved sequent'tr(P* Q) F OP* OQ'
as a derived rule at line 5. This sequent is used, however, in the form'tr(a = b *
la = b) F Oa = b * Ola = b', where'a = b'replaces 'P'and'Ea = b'replaces'Q'.
332 CHAPTER ll
Similarly, the previously proved sequent 'O trP F P' is used in the form 'O la = b F
a=b'atline7.
1
∼a=b A
2
∼□∼a=b H(br∼ I)
3
◇a=b 2 DUAL
4
◇□a=b 3,5-→ E
7
a=b 6◇ □P卜 P
8
--tr-a = b 2-8∼ I
0
tr-a=b 9∼ E
1. ◇P A
2.□ ◇◇P lB
3。 □(◇ ◇P―→ ◇P) N(◇ ◇P→ ◇P)
4。 □◇◇P→ □◇P 3K
5。 □◇P 2,4-→ E
1。 ◇□P A
2. □(◇ □P→ P) N(◇ □P→ P)
3。 □◇□P―→□P 2K
4. □◇□P l◇ Pト ロ◇P
5。 □P 3,4-→ E
Exercise 11.4
Prove the following sequents:
1.卜 a=b…→□a=b
2。 □Pト ロ(PvQ)
3. トロ∼∼P―→□P
4。 ◇ (P&Q)卜 ◇P
5。 □Qト ロ (P→ Q)
LersNrrzrer'r Mooar Locrc 333
334
KRrprearu Moonl Locrc 335
possible to destroy all weapons of war, but it may not (unfortunately) be practi-
cally possible. Logical, metaphysical, physical, and practical possibility are all
forms or degrees of alethic possibility. And there are, no doubt, other forms of
alethic possibility as well. Furthermore there are, as we saw earlier, various non-
alethic forms of "possibility": epistemic possibility, moral permissibility, temporal
possibility, and so on. Does Leibnizian semantics accurately characterize them
all-or do some modalities require a different semantics?
Consider the metatheorem, proved in Section 11.2, that any sequent of the
form tr@ F @ is valid. This seems right for all forms of alethic possibility.'What is
logically or metaphysically or physically or practically necessary is in fact the case.
There are corresponding principles in epistemic, temporal, and deontic logic:
so@
The first is likewise valid. But the temporal and deontic principles are invalid.
'What
was may be no longer, and what ought to be often isn't. Both temporal logic
and deontic logic, then, have non-Leibnizian semantics.
Or, to take a more subtle example, consider sequents of the form trO F !!@,
which are also valid on a Leibnizian semantics. Some variants of this principle in
different modalities are given below:
The temporal and alethic versions are plausible, perhaps; but the epistemic and
deontic versions are dubious. The epistemic version expresses a long-disputed
principle in epistemology; it seems, for example, to rule out unconscious knowl-
edge. And the deontic version expresses a kind of moral absolutism: The fact that
something ought to be the case is not simply a (morally) contingent product of
individual choice or cultural norms, but is itself morally necessary. These are con-
troversial theses.'We should suspect a semantics that validates them.
In fact, Leibnizian semantics seems inadequate even for some forms of alethic
modality. Consider the sequent 'P l- tr O P'with respect to physical possibility. (This
sequent is valid given a Leibnizian semantics; see problem 5 of Exercise 1.1..2.2.)
What does it mean for something to be physically possible or physically
necessary? Presumably, a thing is physically possible if it obeys the laws of physics
and physically necessary if it is required by those laws. But are the laws of physics
the same in all worlds? Many philosophers of science believe that they are just the
regularities that happen to hold in a given world. Thus in a more regular world
there would be more laws of physics, in a less regular world fewer. If so, then the
laws of physics-and physical possibility-are world-relative.2 Leibnizian seman-
tics treats possibility as absolute; all worlds are possible from the point of view of
each. But our present reflections suggest that physical possibility, at least, is world-
relative.
To illustrate, imagine a world, world 2, in which there are more physical
laws than in the actual world, which we shall call world 1. In world 2, not only do
all of our physical laws hold, but in addition it is a law that all planets travel in
circular orbits. (Perhaps some novel force accounts for this.) Now in our universe,
planets move in either elliptical or circular orbits. Thus in world 1 it is physically
possible for planets to move in elliptical orbits (since some do), but in world 2
planets can move only in circular orbits. Since world 2 obeys all the physical laws
of world 1, what happens in world 2, and indeed world 2 itself, is physically
possible relative to world 1. But the converse is not true. Because what happens in
world 1 violates a physical law of world 2 (namely, that planets move only in
circles), world 1 is not possible relative to world 2. Thus the very possibility of
worlds themselves seems to be a world-relative matter!
Kripkean semantics takes the world-relativity of possibility seriously. Within
Kripkean semantics, various patterns of world-relativity correspond to different
logics, and this variability enables the semantics to model a surprising variety of
modal conceptions.
The fundamental notion of Kripkean semantics is the concept of relative
possibility (which is also called alternativeness or accessibility). Relative possibility
is the relation which holds between worlds x and y just in case y is possible relative
to x. The letter '.,?' is customarily used to express this relation in the metatheory.
Thus we write
'z I should confess that virtually everything I am saying here is controversial. But I
have suppressed objections, not because I am confident that what I am saying here
is true, but because I am trying ro rrace a line of thought that makes the transition
from Leibnizian to Kripkean semantics intelligible. The metaphysics I spin out in
the process should be regarded as illustration, not as gospel.
KRrprrnn Mooer Loctc 337
x-Ry
to mean "y is possible relative to ,c" or "y is an alternative to ,r" or "y is accessible
from x." (These are all different ways of saying the same thing.) So in the example
just discussed it is true that 1.n2 ("world 2 is possible relative to world 1"), but it
is not true that 2-e1.. Each world is also possible relative to itself, since each obeys
the laws which hold within it. Hence we have 1.n1. and2.ez.The structure of this
two-world model is represented in the following diagram, where each circle stands
for a world and an arrow indicates that the world it points to is possible relative
to the world it leaves:
{<1,2>,11,1), <2,2>}
The definition of a Kripkean model mimics that of a Leibnizian model, with the
addition of the requirement that.R be defined (item 2 below):
The addition of -rl brings with it a slight but significant change in the valua-
tion rules for '!' and ' O '. Necessity at a world w ts no longer simply truth in all
worlds, but truth in all worlds that are possible relatiue to w. Likewise, possibility
in u is truth in at least one world that is possible relatiue to w.Thus,instead of the
valuation rules 1. 1 and 12 for Leibnizian semantics (Section 11.2), Kripkean se-
mantics has the modified rules:
11' "l/(DO, w) =T iff for all worlds z such that wJtu,Y(Q, u) =l;
V'(!O, w) =F iff for some worldu, w-Ru andY(e,u) +T.
12' \/- ( O @, w) = T iff for some w orld u, w-Ru and Y (@, u) = l;
y( O O, w) =F iff for all worlds z such that u.t.Ru,Y(@, ul *T.
"[('P', 1) = 1
Y(',P"2) =p
as in that example. (That is, planets move in elliptical orbits in world 1 but not in
world 2.) Now the sequent 'P F n O P', which was valid on Leibnizian semantics, is
invalid on this Kripkean model. For V('P', 1) = T, but V('lOP', 1) + T. That is,
world 1 provides a counterexample.
'We
can see that V('l O P', 1 ) + T as
follows. Note first that the only world in
w, accessible from world 2 is 2 itself; in other words, the only
world z in,w, such
that 2..<lt is world 2. Moreover, q{(P',2) * T. Hence for all worlds z in .w, such
that2-&,t,Y('P',u)+T.Sobyrule 1.2',Y(' OP',2) +T.Therefore,since 1-R2,there
is some world x in.l,lz, (namelS world 2) such that 1.tx and V('OP', x) * T. It
follows by rule L1'that "l/('lOP', L) + T. \X/e restate this finding as a formal
metatheorem:
'Pl-n
above.
Moreover, neither of the other sequents mentioned in this section-'trP F P' and
'trP l- tr!P'-is valid, either. Let's take 'lP F P' first.
Knrprraru Nrlooar Loctc 339
This result poses a problem. Intuitively, 'trP I P' is (or ought to be) valid on the
alethic and epistemic interpretations. But it should not come out valid on the
deontic interpretation (which, to distinguish it from the other interpretations, we
usually write as 'OP F P') or on the temporal interpretation discussed above.
The reasoning for the deontic interpretation is straightforward. Think of
world 1 as the actual world, world 2 as a morally perfect world, and 'P' as express-
ing the proposition "Everything is morally perfect." Then, of course, 'P' is true in
world 2 but not in world 1. Moreover, think of =( as expressing the relation of
permissibility or moral possibility. Now world 2 is morally permissible, both rela-
tive to itself and relative to world 1 (because what is morally perfect is surely
morally permissible!). But world 1 is not morally permissible, either relative to
itself or relative to world 2, because all kinds of bad (i.e., morally impermissible)
things go on in it. Our model, then, looks like this:
Now since in this model every world that is morally permissible relative to
the actual world is morally perfect (since there is, in the model, just one such
world, world 2), it follows (by the semantics for 'tr', i.e., formalln rule 11') that it
ought to be the case in world 1 that everything is morally perfect, even though that
is not the case in world 1. Thus, when we interpret'!'as "it ought to be the case
that,"3 we can see how 'nP F P' can be invalid. Kripkean semantics, then, seems
3 Ve could, of course, have used the symbol 'O' instead of 'o' to express the deontic
reading, but we are considering several different readings simultaneously here.
340 CHAPTER 1 2
right for the deontic interpretation, but wrong for the epistemic, temporal, and
alethic interpretations.
But in fact Kripkean semantics is applicable to the other interpretations, as
well, provided that we are willing to relativize our concept of validity. The key to
this new conception can be found by reexamining the proof from an alethic view-
point. From this viewpoint the proof is just wrong. Surely, if it is alethically neces-
sary that P, then P. But where is the mistake?
It lies, from the alethic point of view, in the specification of -4. The alethic
sense of possibility requires that euery world be possible relative to itself, for what
is true in a world is certainly alethically possible in that same world. But the
relation -( used in the proof does not hold between world 1 and itself. The model
is therefore defective from an alethic point of view.
To represent the alethic intepretation, we must insist that -,? be reflexive-
that each world in the set ,rzr. of worlds be possible relative to itself. Thus the
model given above as a counterexample is not legitimate for the alethic interpre-
tation. The only admissible models-the only models that count-for the alethic
interpretation are models whose accessibility relation is reflexiue. This is also true
for the epistemic modalities, but not for the deontic or temporal ones.
This suggests the following strategy: Each of the various modalities is to be
associated with a particular set of admissible models, thar set being defined by
certain restrictions on the relation -r:. ValiditS then, for a sequent expressing a
given modality is the lack of a counterexample among admissible models for the
particular sorts of modal operators it contains. Other semantic notions (consis-
tency, equivalence, and the like) will likewise be defined relative to this set of
admissible models, not the full range of Kripkean models. In this way we can
custom-craft a different semantics for each of the various modalities.
Let us, then, require admissible models for alethic or epistemic modalities,
but not for the deontic or temporal ones, to be reflexive. Then we must redefine
the notion of a valid sequent as follows:
METATHEORE卜 1:All
to the set of models whose
PR00F:
@ there exists a Y whose
KRrprraru Moonr Locrc 341
'We
may say, then, that all sequents of the form trO F O are valid when '!' is
interpreted as an alethic or epistemic operator, but not if we interpret it as a deontic
or temporal operator of the sort indicated earlier. But the validity of all sequents
of this form is the same thing as the validity of the T rule introduced in Section
11.4. Thus we may conclude that the T rule is valid for some modalities but not
for others.
It is the reflexivity of the accessibility relation that guarantees that sequents
of the form trO F O are valid. Such sequents were valid as a matter of course on
Leibnizian semantics, where it is assumed that each world is possible relative to
each, and hence that each world is possible relative to itself. Accessibility in Leib-
nizian semantics is therefore automatically reflexive. But Kripkean semantics li-
censes accessibility relations that do not link each world to each, thus grounding
the construction of logics weaker in various respects than Leibnizian logic.
Just as the reflexivity of .R guarantees the validity of n@ F O, so other require-
ments on -.( correspond to other modal principles. Principles which hold for all
Kripkean models apply to all the logics encompassed by Kripkean semanrics.
Those which hold only in restricted classes of Kripkean models (such as models in
which -( is reflexive) are applicable to some intepretations of the modal operators
but not to others.
We noted above that the principle -O F ntrO seems plausible for temporal
and alethic modalities, but questionable for deontic and epistemic ones. This prin-
ciple is in fact just the 54 rule discussed in Section 1,1,.4.k is valid on Leibnizian
semantics, as we saw in the previous chapter, but it is invalid on Kripkean seman-
tics, since, for example, the instance'!P l- EtrP' is invalid:
Yet the 54 rule is valid relative to models whose accessibility relation is transitive.
The relation .( is transitive if and only if for any worlds x, y, and z, if x.Ry and
y4z, then x-ce. Think of this in relation to physical possibility. \We said that a
world y is physically possible relative to a world x if and only if y obeys the same
physical laws (and perhaps some additional physical laws as well). That is,
x-ey ifand only if y obeys all the physical laws that hold in r.
Now clearly if y obeys all the laws that hold in x and z obeys all the laws that hold
in y, then z obeys all the laws that hold in x. That is, if x-ry and y-Rz, then x-qz.
So the accessibility relation for physical possibility is transitive. The next meta-
theorem shows how we get from this fact about the accessibility relation to the
conclusion that all sequents of the form tr(D l- !trO are valid, where 'l' is interpreted
as physical necessity.
Y(tro, v *
that y.Rz and
y.RZ,
that \/', o ul 1C4u.
ular 1r o, zl to
of t- tro are relative
ls
For our last example, we return to the principle <D l- n O @, which was valid
on Leibnizian semantics (indeed, it is just the B rule introduced in Section 11.4)
but seemed invalid for physical possibility. (The fact that planets move in elliptical
orbits does not mean that it is necessarily possible that planets move in elliptical
KnrprraN lr/ooar Locrc 343
orbits, for there are physically possible worlds in which planetary orbits are nec-
essarily circular and hence in which elliptical orbits are impossible.) The property
of .n that would make this sequent valid is symmerry. .A, is symmetric if and only
if for any worlds x and y, if x-rfii, then y-Rx. The accessibility relation for physical
possibility is not symmetric, since a world with our physical laws plus some "ex-
tra" laws would be physically possible relative to our world, bur ours would not
be physically possible relative to it (since our world violates its "extra" laws).
Logical possibilit5 however, presumably does have a symmetric accessibility
relation-assuming (as is traditional) that the laws of logic are the same for all
worlds. The 6nal metatheorem in this section shows why symmetry guarantees the
validityofOfaOO.
r)*T. Now
we concluded
r) = T, which is a
So, our o relative to the set of
accessibiliry QED
We have said so far that the accessibility relation for all forms of alethic
possibility is reflexive. For physical possibility, I have argued that it is transitive as
well. And for logical possibility it seems also to be symmetric. Thus the accessibil-
ity relation for logical possibility is apparently reflexive, transitive, and symmetric.
It can be proved, though we shall not do so here, that these three characteristics
together define the logic 55, which is characterizedby Leibnizian semantics. That
is, making the accessibility relation reflexive, transitive, and symmetric has the
same effect on the logic as making each world possible relative ro each.
Leibnizian semantics can in fact be viewed as a special case of Kripkean
semantics-the case in which we restrict admissible models to those whose acces-
sibility relation is universal, that is, those in which each world is accessible from
each. Universal relations are, o{ course, automatically reflexive, transitive, and
symmetric. Thus, for example, any sequent which is valid in all reflexive models is
also valid in all universal models. Sequents valid on Leibnizian semantics can from
the Kripkean perspective be regarded as sequents valid relative to the special class
of models with universal accessibility relations. Since Leibnizian semantics seems
3M Crapren 12
appropriate for logical possibility, from a Kripkean point of view logical possibility
is characterized by the class of Kripkean models with universal accessibility
relations.
'If we drop the requirement of symmetry, we lose the law @ l- !O@ (the
inference rule B of Section 11.4), and principles derivable from it, and obtain
a weaker logic, 54, which is a good candidate for being the logic of physical
possibility.
Logics for the other modalities involve other principles and other properties
of -R, many of which are disputed. The chief merit of Kripkean semantics is that it
opens up new ways of conceiving and interrelating issues of time, possibility,
knowledge, obligation, and so on. For each we can imagine a relevant set of worlds
(or moments) and a variety of ways an accessibility relation could structure this
set and define an appropriate logic. This raises intriguing questions that, were it
not for Kripke's work, we never would have dreamed of asking.
Exercise 12.1
Prove the following metatheorems. [Note that saying that a form is valid relative
to the set of all Kripkean models is just another way of saying that it is (unquali-
fiedly) valid on Kripkean semantics.l
1. O F O@ is valid relative to the set of models whose accessibility relation is
reflexive.
2. O O O F O O is valid relative to the set of models whose accessibility relation is
transitive.
3. OEO l- O is valid relative to the set of models whose accessibility relation is
symmetric.
4. 'P I OP' is not valid relative to the set of all Kripkean models.
5. 'OOP l- OP'is not valid relative to the set of all Kripkean models.
6. 'O nP F P' is not valid relative to the set of all Kripkean models.
7. !(@ - Y) i- n(D + !Y is valid relative to the set of all Kripkean models.
8. n<D F 0 O is valid relative to the set of models whose accessibility relation is
reflexive.
9. O O v O -@ is valid relative to the set of models whose accessibility relation is
reflexive.
10. - O (O & -O) is valid relative to the set of all Kripkean models.
In Section 11.4 we introduced the full Leibnizian logic 55. Since then we have seen
that some of the rules of S5 are inappropriate for certain forms of modality. The T
rule (from l<D infer @), for example, is plainly invalid when 'n' is taken to express
obligation, as it is in deontic logics. We have now seen that this rule was validated
by the reflexivity of the accessibility relation. Likewise, the B rule (from O infer
KRrprraru l\zlooar Locrc 345
Duality (DUAL) From either of O@ and -r-@, infer the other; from either
of ilD and - O -@, infer the other.
K rule (K) From l(@ * Y), infer (n(D * nY).
Necessitation (N) If O has previously been proved as a theorem, then any
formula of the form !(D may be introduced at any line of a proof.
These rules are common to all Kripkean modal logics. Together with the ten basic
rules of classical propositional logic they constitute a logic that is sound and com-
plete relative to the set of all Kripkean models. This logic is known as the system
K (for Kripke!). In other words, a sequent of propositional modal logic (modal
logic without the identity predicate or quantifiers) is provable in the system K iff it
has no counterexample in any Kripkean model.a
K itself is not very interesting. But by adding various rules to K we may
obtain differing logics that are useful for different purposes. Each rule corresponds
to a particular structural requirement on the accessibility relation. Imposing new
structural requirements diminishes the range of admissible models-models that
may serve as counterexamples. Thus imposing new structural requirements on -,?
increases the number of valid rules. Among systems we have considered, the one
with the most structural requirements is 55, for whose admissible models -.( must
be reflexive, transitive, and symmetric. In a sense 55 is the maximal Kripkean logic,
since it is sound and complete for the most restrictive class of models, the class of
models whose accessibility relation is universal. (Though reflexivity, transitivity,
and symmetry don't entail universality, the class of all universal models determines
the same logic, 55, as the class of reflexive, transitive, and symmetric models does.)
Exercise 1 2.2
Note thatin the problcrns bclow itis not safe to usc thc scqucnts proved in Section
ll.4 as derived rulcs,since these were prOvcd using thc full logic of S5 arld the
systems in which we are wOrking are■ agments of SS in which certain rules are
unavailablc.Ncvcrtheless,sorne of thc stratcgics illustrated in that section inay be
useful herc.
have until now been using the material conditional, symbolized by *', to
'We
render the English operator 'if . . . then' into formal logic. This practice, as we
noted in Section 3.1, is, strictly speaking, illegitimate. The material conditional is
at best only a loose approximation to 'if . . . then'. Many inferences which are
valid for the material conditional are invalid for English conditionals. Consider,
for example:
Socrates grew to manhood.
If Socrates diedas a child, then Socrates grew to manhood.
TABLE 12.1
Some lmportant Kripkean Propositional Modal Logics
Accessibility
Logic Basic Rules* Relation Application
It is not the case that if Socrates was a rock then Socrates was a man.
Socrates was a rock, but not a man.
In each case, the premises are true and the conclusion is false in the actual world,
using our ordinary understanding of the conditional. Yet in each case, the argu-
ment is valid if we interpret 'if . . . then' as the material conditional. The last two
arguments have forms that at first glance appear to be paradigms of good reason-
ing: hypothetical syl logism,
A*B,B.+CFA-C
and contraposition,
A*Bl--B*-A
(sometimes called "transposition"). Yet these forms are apparently invalid for
'if...then'.
C. I. Lewis, the inventor of 54, 55, and other modern modal systems, was
one of the first formal logicians to investigate the disparity between English and
material conditionals. Lewis noticed that ordinary English conditionals seemed to
express, not just a truth function, but a necessary connection between antecedent
and consequent. Defying skepticism about the intelligibility of the concept of nec-
essary connection, Lewis introduced in 1918 a new conditional, represented by
the symbol '3', which incorporated this idea. O 3 Y is true if and only if it is
impossible for both <D to be true and Y false. O 3 Y is true, in other words, if and
only if it is necessarily the case that if (D then Y, where 'if . . . then' signifies the
material conditional. Thus '3' is often introduced as a defined operator into modal
systems using the definition
(D3Yiffnlo-Y;
An equivalent definition in terms of the possibility operator is
(D3Yiff-o(o&-Y)
Translated into Kripkean semantics, the truth conditions for the strict conditional
are as follows:
y(o 3 Y, w) = T iff for all worlds z such that w-Ru andY(@,u) =
T, Y(Y, u) =T
y(O 3 Y, w) - F iff for some world er such that w.Ru,Y(@,u) =l
and 1/(Y, u) +T
\We began this section with five arguments, the forms of the first three of
which were as follows:
BFA*B
-Al-A*B
-(A.*B)FA&-B
All three arguments are valid, reading'*' as the material conditional, but all are
outrageous reading '*' as an English conditional. (Indeed, the first two have often
been called the "paradoxes of material implication.") Yet, if we replace '-'
by '3',
we get the reasonable result that none of the three arguments is valid.
Let's consider the sequent 'B F A 3 B' first. To facilitate comparison with the
first argument above, think of 'B' as meaning "socrates grew to manhood" and'lf
as meaning "socrates died as a child." Socrates did, of course, grow to manhood;
yet it is (or was) possible for him to have died as a child and not grown to be a
man. So the premise is true and the conclusion false. To represent this counter-
example formally we need two worlds: world 1, representing the actual world, a
world in which Socrates did grow to be a man, and a merely possible world, world
2, in which he died as a child:
‐
‐rレ 争=[1,2} γ(`札 1)≡ F
│‐
.γ
?〒 {く 1,1>,<1,2',<2,2ゝ } (`B';1)=T
ν(官 ,2)=T
`′
γ(`B',2)≡ F
But the situation is not as tidy as it seems. The last two of our five arguments
have the forms hypothetical syllogism and contraposition, respectively. These
forms, as we saw, seem invalid for English conditionals, but they are valid for the
strict conditional. We shall prove this for contraposition onlS leaving the proof
for hypothetical syllogism as an exercise:
PRooF: Suppose for reductio that this sequent is invalid relative to the
have a contradiction.
Therefore the sequent 'A € B F *B 3 -A is valid relative to the admis-
sible models for strict conditionals. QED
The fact that it makes hypothetical syllogism and contraposition valid might
be seen as an advantage, rather than a disadvantage ofthe strict conditional. These
are, after all, common and persuasive forms of reasoning. But since they are ap-
parently invalid for at least some English conditionals, their validity for the strict
conditional is in fact a disadvantage, insofar as the strict conditional is supposed
to accurately analyze the English.
The disparity beween strict and English conditionals also crops up in "para-
doxes" reminiscent of the paradoxes of material implication. These concern the
sequents 'trB l- A 3 B' and '-OA F A 3 B', both of which are "paradoxically"
valid. Reading '3' as an English conditional, we can produce preposterously in-
valid instances. For example:
Exercise 12.3
Prove the following metatheorems for the logic of strict conditionals-whose ad-
missible models are all Kripkean models in which -,? is reflexive.
1.. The sequent'-A l- A 3 B' is invalid.
2. The sequent'-(A 3 B) F A & -B' is invalid.
3. The sequent'trB l- A 3 B' is valid.
4. The sequent '- O A F A 3 B' is valid.
5. The formula 'A 3 A' is valid.
6. The sequent'A 3 B, A F B' is valid.
7. The sequent 'A 3 B, -B l- -1t' is valid.
8. The sequent'A 3 B, B 3 C l- A 3 C' is valid.
9. The sequent'A, B l- A 3 B' is invalid.
10. The sequent'A, -B F -(A 3 B)' is valid.
Contrast this with similarly stated truth conditions for the strict conditional:
O 3Y is true in a world w iff in all the worlds possible relative to u in
which (D is true, Y is also true.
Here, of course, we have to specify the relevant sense of possibility; that is, we
have to know which form of alethic modality we are dealing with.
Lewis's truth conditions, however, do not require us to specify the sort of
possibility we intend. The antecedent of the conditional does that automatically.
'We
are to consider, not all practicallS or physically, or logically possible worlds,
but rather all the worlds most like ours in which the antecedent is true.
As a result, Lewis's truth conditions do not flounder, as those for the strict
conditional do, when the antecedent is impossible. With the strict conditional, if
there are no possible worlds in which the antecedenr is rrue, then, trivially, the
consequent is true in all such worlds-no matter what that consequent may say.
Thus, as we saw, given that it is impossible for Socrates to be a rock and reading
'if . . . then' as a strict conditional using the practical sense of possibility, we must
concede that the absurd sentence 'lf Socrates is a rock, then Socrates is a chi-
huahua' is true.
Lewis's semantics avoids this consequence. Having not found any practically
possible worlds in which the antecedent is true, we do not simply punt and declare
the conditional true; rather, rising to the challenge, we consider more and more
remote possibilities. In our example, since it seems impossible, even in the meta-
physical sense, for Socrates to be a rock, we must extend our consideration all the
way out to mere logical possibilities before finding worlds in which he is. lfhen
we come to the first of these (i.e., those most like the actual world-so that, despite
the fact that in them Socrates is a rock, as much as possible of the rest of the world
is as it actually is), we stop. Then we ask: Is Socrates a chihuahua in all of these
worlds? The answer, pretty clearly, is no. And so the sentence 'If Socrates is a rock,
then Socrates is a chihuahua' is false.
Though this example is artificial, the general procedure is not. When consid-
ering whether or not a statemenr of the form if @ then Y is true, we do in fact
imagine things rearranged so rhat (D is true and then try to determine whether
under these new conditions Y would also be true. But we do this conservatively,
excluding ways of making @ true that are wilder than necessary. That is, we try to
keep as much as possible of our world unchanged. Most of us would assent to the
conditional 'if kangaroos had no tails, they would topple over', even though we
can envision worlds in which kangaroos have no tails but do not topple over
because, for example, there is no gravity. But the conditional asks us only to
entertain the possibility of depriving kangaroos of tails. Depriving them of gravity
too is impertinent; it changes the world in ways not called for by the conditional's
antecedent. Hence, depriving kangaroos of gravity is not relevant to determining
the truth value of the conditional.
Yet there may be more than one equally conservarive way of changing the
world to make the antecedent true. Consider the conditional 'if forests were not
green, then they would not be so beautiful.' Now there are many worlds equally
minimally different from ours in which rhe antecedent is true: worlds in which
forests are brown or blue or yellow, and so on. Only if we regarded the consequent
as true in all of these worlds would we assent to the conditional. If we regarded
brown forests, but not blue, as more beautiful than green, we would judge the
conditional false. That's why Lewis stipulates that if @ thenY is true iff among all
the worlds (plural) most like ez in which @ is true, Y is also true.
The one element required by Lewis's semantics that has not appeared in any
model we have considered so far is a measure of "closeness" or similarity among
worlds. \7hile Lewis uses these terms, I prefer to think in terms of degree of
possibility; where Lewis would speak of worlds as being more or less similar to a
given world, I regard them as being more or less possible relative to that world.
Knrprran lVooar Locrc 353
There are two reasons for this. First, it allows us to make the transition to Lewis's
semantics without introducing the entirely new concept of similarity; the only
change we need make is to think of -F as having degrees, rather than being an all-
or-nothing affair. Second, similarity is symmetric; A is precisely as similar to B as
B is to A. But, as we have seen, =( should not, in general, be assumed to be
symmetric.
How might a model treat Jl as a matter of degree? The simplest way would
be to set up some arbitrary scale (say, from 0 to 1), where 0 represents complete
lack of relative possibility and 1 the highest degree of relative possibility. Presum-
ably, then, each world is maximally possible relative to itself, that is, has degree 1
of -R to itself, and all other worlds are less possible relative to it.
Such a numerical scale is, however, not quite satisfactory. There is no a priori
reason to suppose that degrees of relative possibility can be ordered like the real
numbers from 0 to 1. A more abstract mathematical treatment of order could
address this problem but would introduce complexities beyond the scope of this
book. \We shall, then, at the risk of slight (and not very significant) oversimplifica-
tion, suppose degrees of -P can be ranked along a 0 to 1 scale.
Accordingly, instead of treating -R as a set of pairs, as we did before, we may
treat it as a set of triples, in which the third member is a number from 0 to 1,
indicating the degree to which the second member is possible relative to the third.
Thus for a model consisting of worlds 1. and 2, we might have, for example:
and that
so that for any worlds x and y in the model ,.R(x, y) will exist and will be unique.
(I7here in a Kripkean model we would say that it is not the case that x.Ry, now
'We
we shall say .R(x, y) = 0.) further stipulate that
(3) for any worlds x and y, -plx, y) = 1 iff x = y.
This implies that no world is as possible relative to a world x as r itself is. A Lewis
model, then, will be exactly like a Kripkean model except for these differences
in.A.
Lewis represents his conditional formally as the binary operator '5-', and
we shall do likewise. But we shall differ from Lewis in reading this operator simply
as "if . . . then." Lewis reads <D l- y as "if (D were the case, Y would be the case,"
confining his analysis to so-called subjunctive or counterfactual conditionals. But
354 CHAPTER 12
I am persuaded that this analysis is more broadly applicable.5 Its formal truth
clause is as follows:
V((D n-Y, wl :T
iff there is some world, u such that Y(@, u) = T, and there
is no world e such that .R(w, z) 2 -e(u u),Y(@, z) =T, and V(Y, z) +T.
This is just a transcription in our new terminology of the informal truth conditions
given above. The world z is some arbitrary one of the worlds most possible relative
to the actual world in which the antecedent (D is true. 'We are saying, in other
words, that @ tr+ Y is true at w iff.
1. @ is true in some world z, which is such that
2. there is no world at least as possible relative to w as z is in which @ is
true and Y is not.
Clause 2 implies that Y is true in w, as well as in any worlds more possible relative
to w in which @ is true. Putting both clauses together, this is to say that in all the
worlds most possible relative to w in which O is true, Y is also true. The corre-
sponding falsity clause is
V((D l* Y, w) = F iff for all worlds z such that "l/(<D, u) =T rhere is some
world z such that .R(a z) 2 -R(w, u),Y(Q, z) =T, and 7(Y, z) +T.
If we wish to retain the operators 'O' and 'n', we can do so in the Kripkean
fashion, by stipulating that for any worlds x and y, x.ny iff -n(x, y) + 0. That is, y
counts as accessible from x if and only ify is accessible to even the slightest degree
from r. This allows the standard Kripkean clauses to be used for these operators.
We shall illustrate the use of Lewis semantics first by proving that modus
ponens is valid for a Lewis conditional:
defined as follows:
oも 二
`И
ー
マ■‐{│く 1,1,1>,く 1,2, 0。 7>,<2,1, ・
:│││17i(BI:‐ 1,‐ =T .││
l γ(`″ :12,■ TIIIIII I
│‐
“
│││や
1,す 1>l ‐
│
│ │││││││■ ■│
│││││■
γ(`B'12)=F ‐││■
` ′
│‐
_│││. .′ │■
1-3 and
355 Craprrn 1 2
Exercise 12.4
Prove the following metatheorems for 'tr-' using Lewis models.
1. The sequent'A l- B, -B l- -1t' is valid.
2. The sequent'A l- B, B a- C f A l* C'is invalid.
3. The sequent'-A l- A tr+ B' is invalid.
4. The sequent'B F A E- B' is invalid.
5. The sequent'-(A n* B) F A & -B' is invalid.
6. The sequent'A, B F A n- B' is valid.
7. The sequent'A, -B l- -(A l* B)' is valid.
8. The sequent'A D* C F (A 6c B)l* C' is invalid.
9. The sequent'!B l- A tr* B' is invalid.
10. The sequent '- O A F A l* B' is invalid.
DEONTIC AND TENSE LOGICS
C月 ス PTfR ,3
Deontic logic has often been studied as it was presented (briefly) in Chapter 12-
independently of alethic logic, using Kripke-style models whose accessibility rela-
tions represent permissibility rather than alethic possibility. Though this approach
has been fruitful, more can be learned by considering deontic and alethic operators
together in the same model. Some important deontic principles, such as Immanuel
Kant's dictum that ought implies can, involve both deontic and alethic elements
(the 'can' signifies some form of alethic possibility). These principles remain unex-
amined when deontic logic is studied in isolation.
To do modal and deontic logic together, we need two pairs of monadic modal
operators:'tr' and'O' to express necessity and possibility, and'O' and'P' to
represent obligation and permission. Thus we also need two accessibility relations:
an alethic relation, which we represent, as before, by the letter '.4', and a deontic
relation, which we represent by the letter 'J'. For any two worlds x and y, x.ny if
and only if y is possible relative to x, and xsy if and only if y is morally permissible
relative to x.
'u7hat
it means for y to be morally permissible relative to x is a question that
lies outside logic. Presumably, for a utilitarian like Jeremy Bentham or John Stuart
Mill it means that, relative to the conditions that hold in x (the kinds of sentient
357
358 CuaprEn 13
beings, desires, and so on that exist there), world y is a world in which happi-
ness is maximized. For a deontologist like Kant it might mean that in y all actions
accord with the moral law that holds in x-and perhaps for Kantians the same
moral law holds in all worlds so that any world that is permissible relative to
one world is permissible relative to them all. The logic to be presented here is
compatible with these interpretations, as well as others, and does not decide be-
tween them.
The valuation rules are just like those for Kripkean modal logic except that
we need two new clauses for the deontic operators. These mirror the clauses for'
l' and ' O ', except that they have s in place of -rl:
13. y(OO, w) =T iff for all worlds z such that wsu,Y(@,u) =l;
y(OO, w) =F iff for some world u, wau andY(Q, u) +T.
14. y(PO, w) =T iff for some world u, wsu andY(@,u) =\
y(PO, w)=F iff for all worlds z such thatusu,Y(@,u\+T.
PR00F:
such that it is not
consider z such
w51,!, such that
wl T.
Dronrrc aruo Te rusr Locrcs 359
Thus we have shown that if s is not serial, then there is a world aa such
that Y(OO, w) =T for any formula <D. QED
│‐
│‐
││‐ │‐
‐ │.',‐
‐ヽ
│‐
11,‐
:‐ 11‐
│‐
`.‐
│‐
.‐
│
'│.‐
・
tl‐
Suppose, first, that 1r(OO, w) =T and "l/(-P-@, wl +T,
Since V(O(D, La) = T, it follows by rule 13 that for all
│..│‐ │‐
'We
also, of course, have the dual result:
│ │■ ││■ │■ ■│
This model is admissible, since -^, is reflexive and s is sleriali S五 こ
ё12 is
the only world u in this model such that lsw and since Y('1t', 2l =T,it
follows that Y('1t', u) =T for all worlds z such that'1.su. Thus by rule
13, "l/('O/t', 1) = T. But since 1 is the only world ez such that 1"4u, and
since Y1'A', I ) + I there is no world a such that latu and'l/1'.A', 1) = T
Hence by rule 1,2,1r( OA', 1) * T. But then since both Y('OA', 1) = T
and V('OA', 1)* T, by rule7, y('OA
Olf is not valid. QED
- OA', 1)+ T; and so
The notion that ought implies can would seem ro be grounded more in the warm
cosmic optimism of the Age of Enlightenment than in cold logic. But the whole
story is not yet told; it takes a new twist at the end of this section.
Meanwhile, let's consider a more practical application. Unsophisticated rea-
soners often argue in the following pattern:
DroNrrc aro TrNse Locrcs 361
OB,A*BFOA
For instance:
'We
ought to eliminate AIDS. If we adopt a policy of strict quarantine of
AIDS victims, we eliminate AIDS. Therefore, we ought to adopt a policy of
strict quarantine of AIDS victims.
With respect to this sequent, 'we eliminate AIDS' is 'B'and'we adopt a policy of
strict quarantine of AIDS victims' is '1t'. The fallacy is obvious: There may be
better policies that would achieve the same goal-like finding a cure. That is, we
can envision a world, call it u, in which AIDS ought to be eliminated (i.e., is
eliminated in all of eo,'s morally permissible alternatives) and in which a quarantine
would eliminate AIDS, but in which it is not the case that we ought to quarantine
(i.e., there is a morally permissible alternative world in which we do not quaran-
tine). I7e can model this counterexample with just two worlds-world 1, which
we may think of as the actual world, and world 2, a possible and permissible
alternative to world 1 in which AIDS is eliminated by some means other than
quarantine. Here is what the model looks like in formal terms:
F F F T
〓 〓 〓 〓
wu= 11,21 v(
1
F
1/(
Since I
V('OB', tu) * by rule 13 there is a world z such that
wsu and Y ('B', u) + T. But since V('O1t', w) = T and wsu, it
follows, again by rule 13, that Y('A, u) =T. And since lf('y'f,
u) =T and V1'B', u) *T, it follows by rule 7 that Y('(A *
B)',
u) + T. FinallS since wau, it follows again by rule 1,3 that
?"('O(A *
B)', wl *T. But we said above that V('O(A r
B)',
w) =T, and so we have a contradiction.
Thus, contrary to our hypothesis, the sequent'OA, O(A *
B) F OB'is
valid. QED
It's now time to take up some business left unfinished at the end of Section
9.4. This concerns the argument
I should live.
It is necessarily the case that if I live Bad Bart dies.
Bad Bart should die.
'We
saw in our earlier discussion that the form of this argument is not straightfor-
wardly representable in predicate logic. But it is easily representable in modal
deontic logic. Using'11 for 'l live' and 'B' for'Bad Bart dies', the form is:
OA, tr(A * B) F OB
This is invalid. For, as in the case of ought implies can, maybe what we have here
is a genuine moral tragedy. Perhaps what ought to be the case is that no life is lost,
even though, given the circumstances, this can't be the case. 'We can thus use a
counterexample similar to the one used to refute the thesis that ought implies can.
Once again, world 1 is the actual world-in which, we shall suppose, Bart gets
me. (I hesitated too long in my moral deliberations.) World 2 represents what
ought to be-a world in which we both live, but which, unfortunately, is impossi-
ble relative to world 1.
This result, however, wouldn't have helped Kant, for he clearly intended the 'can'
in his principle 'ought implies can' to express practical possibility.
As the following exercise illustrates, the valtdity 'ought implies can' is only
one of many changes that our logic undergoes if we move from our former notion
of admissible models to universal models-that is, from physical or practical ro
logical possibility.
Inferentially, our modal deontic logic requires separate rules for the modal
and deontic operators. The rules DUAL, K, and N (see Section 11.4), being valid
for all Kripkean logics, are valid for both sets of operators. Since we have required
s to be serial, we have the rule D (from OtD infer P@) for the deontic operarors
(the proof of its validity is left as an exercise below). If we merely require -rf to be
reflexive, we need for the modal operators only the additional rule T, but if we
require =,? to be universal, we must employ the full Leibnizian logic, including I
B, and 54. These rules, then, constitute logics that are sound and complete relative
to the semantics presented here.
Exercise 13.1.1
For each of tlle following scqucnts,prove nletatheorctically cithcr that it is valid
or that it is invalid,rclative to the class of adnlissiblc nlodcls in、 vhich ιマis rcncxive
andざ is scrial.Whcrc thc sequentis invalid,dcscribc a cOunterexanlplc infOrinally,
as、 vcllas giving thc fornlal pr00f.I)Oes the validity Of any inference change if we
requircこ ″to bc univcrsal in adnlissible n10dclsP
l.A卜 OA
2.OA卜 A
3.OA卜 PA
4.□ A卜 PA
5。 □A卜 10A
6。P(A&B)卜 PA
7.OA,OB卜 0(A&B)
8。 PA,PB卜 P(A8(B)
9。 OA,A― →B卜 OB
10。 A,0(A→ B)卜 OB
ll.A,(A― →OB)卜 OB
12.PA,0(A― →B)卜 PB
13.PA,P(A― →B)卜 PB
14。 PA,□ (A― →B)卜 PB
DeoNrrc aruo Trnsr Locrcs 365
Exercise 13.1.2
Prove that the following formulas are valid relative to the class of models in which
-.( is reflexive and s serial:
1.∼ P(A&∼ A)
2.O(Av∼ A)
3.∼ (OA&0∼ A)
4.∼ (□ A〔に□∼A)
5。 PA∨ P∼ A
"'What then is time?" asks St. Augustine. "I know what it is if no one asks me what
it is; but if I want to explain it so someone who has asked me, I find that I do not
know."2
Ordinarily we understand time as a linearly ordered sequence of moments.
'We have position in time, the present moment. All other moments lie in either
a
the past or the future. The present constantly advances toward the future, and this
advance gives time a direction. The past is a continuum of moments stretching
behind us, perhaps to infinity. It is unalterable. Whatever has been is now neces-
sarily so. The future, however, is not frozen into unalterability but alive with
possibilities. Starting with the present, events could take various alternative
courses. There is, in other words, more than one possible future. Though only one
of these courses of events will in fact be realized (we may not, of course, know
which one), still the others are genuinely possible, in a way that alternative pasts
are not genuinely possible.
These intuitions suggest a model on which time is like a tree with a single
trunk (the past) that at a certain point (the present) begins to split and split again
into ramifying branches (various possible futures). As time moves forward, the
lower branches (formerly live possibilities, lost through passage of time) disappear.
Only one path through the tree represents the actual course of time, that is, the
actual world. More and more of the path is revealed as time moves on and lower
branches vanish. If time were finite, eventually all the branches representing merely
possible futures would disappear and only this single path from trunk to branch
tip would remain: the entire history of the actual world from the beginning to the
end of time. But we might also think of time as infinite-at least toward the future
and perhaps also backward into the past. If time is infinite toward the past, then
the tree's trunk extends endlessly downward, never touching ground; and if time
is infinite toward the future, then its branches stretch endlessly upward, never
touching the sky. In either case, we might picture at least a part of the tree like the
diagram in Figure 13.1.
FIGURE 13.1
A Picture of Time
Future
Present
Past
It would be rash to claim that this picture of time is true. Time, according to
relativity physics, is inseparable from space and relative to motion, and though it
is experienced as past, present, and future, these may not be "objective" features
of time itself. Even if they are, our picture may well be inaccurate for the very
distant past (in the vicinity of the big bang) and for the very distant future, because
we do not fully understand the behavior of time at these extremes. Moreover, if
the world is deterministic, then possible alternative futures (the thin branches) are
merely illusions. But this is not the place for a detailed physical or philosophical
discussion of the nature of time. The logic presented in this section should be
regarded not as the one true tense logic, but as an example of a tense logic that
accords reasonably well with our everyday understanding of time. The literature
on tense logic is replete with logics for relativistic time, discrete time, circular time,
Drorurrc aruo Trrusr Locrcs 357
ending time, and so on, all of which differ substantially from the logic presented
here.3
A tense logic is a logic that includes operators expressing tense modifications.
The logic we shall consider here is a modala tense logic because it contains alethic
modal operators in addition to tense operators. It has four tense operators:
H(D*-P-O
PO*-H-O
G(D*-F-O
F@*-G-<D
Using these four operators, we can express a great variety of tense modifica-
tions. Here are some examples:
Group I
GHO It will always be that it has always been that @. [Intuitively, this
means that (D is the case at all times-past, present, and future.l
FHO It will be the case that it has always been that <D. [O has always
been the case and will continue to be for some time.]
H@ It has always been the case that (0.
PHO It was the case that it had always been that @. [There was a time
before which it was always the case that iD.]
HPO It has always been that it had (at some time) been the case that O,
[That is, there have always been times past at which @ was the
case, but these may have occurred intermittently.]
PO It was the case that O.
GPO It will always have been that <D.
FPO It will be the case that it has (at some time) been that @.
3 The classic source for many of these variants is Arthur Prior, Past, Present and
Future (Oxford: Oxford University Press,1967). Prior's use of Polish notation
may be a barrier for readers accustomed to the more unusual notation employed
in this book, but with a little practice one can easily adapt to it. Many variants of
tense logic are also discussed in Nicholas Rescher and Alasdair Urquhart, Tempo'
ral Logic (New York: Springer-Verlag, 197 ll.
a 'Modal'here is used in its narrow sense, as a synonym for'alethic'.
368 CHaprrn 13
Group II
HGO It has always been the case that it would always be that @. [<D is
the case at all times-past, present, and future.]
PGO It was (at some time) the case that it always would be that (D.
GΦ It always will be the case that @.
FGO It will (at some time) be the case that it will always be that @.
[That is, there will come a time after which @ is always the case.]
GFO It will always be the case that it will sometimes be the case that (D.
[Moments at which @ is the case will always lie in the future,
though perhaps intermittently.l
FO It will be the case that <D.
HFO It has always been that it will be the case that @.
PFO It was (at some time) the case that it would (later) be the case that
o.
An adequate tense logic will, for example, determine which of these statement
forms imply which others, but of course it will do a good deal more than that. (In
fact, on the assumption that there is no first or last moment of time-an assump-
tion discussed below-each formula of either group implies all formulas listed
below it in the same group; and the first members of the tlvo groups are equivalent,
as are the last.)
There is in most tense logics no special operator for the present tense. To
assert that (D is presently the case, we simply assert O. We anticipate, then, that
sequents such as the following are valid:
FHAFA
AFHFA
PGAFA
AFGPA
Our intuitive picture of time includes multiple possible futures. Indeed, each
path through the tree from the base of the trunk (if it has a base) to the tip of a
branch (if branches have tips) represents a complete possible world. These possible
worlds share a portion of their histories with the actual world but split off at some
specific time. This is a picture we often use in decision making. Suppose I am
considering whether to go to the mountains for a hike or iust stay at home and
relax this weekend. These are (we assume) realpossibilities, though undoubtedly
not the only ones. Corresponding to each is at least one possible world-that is,
at least one course of events that the world might take from the beginning of time
through and beyond the moment of my decision. Suppose I decide to hike and I
carry out that intention. Then the world (or one of the worlds) in which I hike is
the actual world, and the worlds in which I stay at home that weekend are possible
but nonactual. In these nonactual worlds, everything up to the moment of my
decision occurs exactly as it does in the actual world, though events depart from
their actual course more or less dramatically thereafter.
Dronrrc nNo Teruse Locrcs 359
lX/hen the weekend is over, I may say, "Though I could have stayed home, I
can't now go back and change the past; it is now necessarily the case that I went
hiking," mixing tenses and alethic modalities in ways that our picture nicely illus-
trates. To say that I could have stayed home is to say that up to the beginning of
the weekend a world in which I stayed home (represented by a path up the trunk
through one of the thin branching lines) was possible. This branch, however, has
disappeared as time has moved on. To say that my having gone hiking is now
necessary is to say that I did go hiking in all currently possible worlds, a circum-
stance represented in our picture by the fact that all currently possible worlds have
exactly the same past as the actual world (the tree has but one trunk).
'What
we have been thinking of as the disappearance of the tree's lower
branches can also be understood in Kripkean terms as the termination of accessi-
bility. In a sense these "vanished" branches are still there; they still representworlds
that are possible in some absolute sense. But these worlds are no longer possible
relative to (i.e., no longer accessible from) the actual world. In tense logic, in other
words, accessibility is time-relative. Thus to represent alethic modalities in familiar
Kripkean fashion in the context of tense logic, we must add a temporal index to
the accessibility relation -(. Instead of saying flatly that world w.is accessible from
world wb we must specify a time relative to which accessibility is asserted. Thus
we shall write, for example, 'we/|tozt'to indicate that w, is accessible from w, at
time /.5 Worlds in which I stayed home on the weekend in question are accessible
from the actual world prior to my leaving, but not thereafter.
Truth, already relativized to worlds in alethic modal logic, must in tense logic
be further relativized to times. It is true now that I am sitting at my computer, but
this will not be true a few hours hence. Thus the statement 'I am sitting at my
computer' is true at one time and not at another within the actual world. More-
over, though it is true now in the actual world, it is not true in a world (possible
until very recently) in which I got up and went for a snack a moment ago. Thus a
statement may have different truth values at different times within the same world
and different truth values at the same time within different worlds. Valuations for
predicates (including zero-place predicates) must, accordingly, be indexed to both
worlds and times. \We shall rvrite, for example,'V(<D, t, w) =T'to indicate that
formula @ is true at time r in world w.But we shall treat names, as before, as rigid
designators, relativizing their denotations neither to worlds nor to times.
Now times (or, as many authors prefer, "moments") do not just occur ran-
domly within worlds, but successively in a strict linear order. In fact, a world may
simply be defined as a linear progression of times. To do this formally, we must
specify the relation by which the times are ordered. We shall call this the earlier
than relation and represent it as 'r'. Thus '/rETr' means that time /, is earlier than
time /r. To say that the times comprising a world are linearly ordered is to say that
for any times ,r and /, belonging to the same world, either trtt, or t2ttt or tt = t2.
This implies that the moments comprising a given world can all be arrayed, as in
our intuitive picture, as points along a single (possibly curved but more or less
vertical) line, with each earlier moment beneath all later moments.
,; moreover, must in general be transitive-that is, for any times /,, tr, and
tb if t et2 and trtt., then trtt.-for it violates our conception "earlier" to think of r,
as earlier than /, and /, as earlier than /. but not f, as earlier than /r.6
Finally, we must recognize that even domains, which in alethic logic were
relativized to worlds, must now be relativized to times as well. Objects come inro
and go out of existence as time passes. Thus within a single world what exists at
one time differs from what exists at another. But also at a given time what exists
in one world may differ from what exists in another. I am now poised over a soap
bubble, ready to pop it with my finger. If I choose to do so, then a moment
afterward the actual world contains one less soap bubble than exists at the very
same moment in the world that would have been actual had I not poked.
Having made these observations, we are now ready to say what a formal
model for modal tense logic looks like:
6 We might also be tempted to insist that r is irreflexive (i.e., for all times t, it is not
the case that ttt) and asymmetric (i.e., for all times t, and tr, if r,:t, then not t2etl).
But we will resist these temptations. Making: irreflexive and asymmetric wouId
prevent us from modeling the kind of circular time envisioned in Nietzsche's idea
of the eternal recurrence and from envisioning other interesting kinks in time, and
there seems to be no a priori reason to rule such models out.
7 This definition permits us to wrire the more intuitive expression 't is in w' rather
than the formal'te t-'. There is no point in trying ro remember the meanings of
too many symbols.
Dronrrc aruo TerusE Locrcs 371
Notice that this definition allows worlds to share times, though it permits
what is true at a time in a given world to be false at that same time in a different
world (as happens after worlds diverge). A world need not contain all times. One
world may begin or end sooner than another, and some might be temporally
infinite-having no beginning and no end. And worlds need not share any times
at all. There may be worlds with times wholly distinct from ours. However, the
earlier-than relation'transcends worlds in the sense that if tft2,t1is earlier thant.,
in any world that includes both of these times. Our definition even allows us to
model worlds that skip times-jumping as it were from an earlier time to one
much later. I doubt that we have much use for such models, but it is harmless to
suppose that logic does not rule them out.
Notice, finally, that our definition of a model does not include a specification
of the alethic accessibility relation -{'. This is because -c is definable in terms
already available to us. Specifically, we may say that world ra, is accessible from
wt at time t iff wrhas exactly the same history as wt op to time t.8 If wrdiffers in
any respect from w, before /, then w, is no longer possible relative to wr, for
otherwise the past would not be necessary. (Note, however, that in order to be
accessible from w, at time t, w, need not diverge from w, precisely at time /; the
divergence of the two worlds may yet lie some distance into the future.) Two
worlds w, and wrhave the same history iff they consist of the same moments up
to time / and every atomic formula that is true at a given moment before / in one
is true at the same moment in the other, that is, if they meet conditions 1-3 of the
following definition:
8 We assume here that if two worlds that are exactly alike up to a certain time , are
both possible (i.e., members of the set of possible worlds), then nothing that oc-
'curs
in either world before that time dictates which of their two courses events
must follow from , onward. Hence at , either world is possible relative to the
other. This assumption seems reasonable, but as we shall see below it implies a
very strong alethic logic, 55. Alternatively, we might have insisted that having the
same initial history up to , is a necessary, though not sufficient, condition for ac-
cessibility at r. This would permit a variety of weaker alethic logics.
372 CHAPTER 1 3
2. w, and w, contain the same times up to /; that is, for all times /', if
tt, then t' is in wriff r' is in wr, and
t'
3. w, and w, have the same atomic truths at every moment up to ,,.
that is, for all times /' such that. t'Et, :l(t', wr) = b(/', wr), and for all
predicates O, Y(O, t', w r) =Y (@, t', wr).
Since names are not relativized to either worlds or times, stipulating that all times
up to , have the same domains and give predicates the same extensions in both
worlds insures that the truth values of atomic formulas are the same in both worlds
up to ,. It also guarantees that most nonatomic formulas have the same truth
values in w, and wr-but not that all nonatomic formulas do. We should expect,
for example that a formula of the form F@ might have different truth values in ru,
and w, even at times before f, since though the two worlds' histories up to , are
the same, their futures need not be.
Notice that we require for accessibility only that a world share its history
with a given world up to the time I in question. Thus, for example, a world may
be accessible from the actual world, even if it differs at the present moment, pro-
vided only that its past history is exactly the same. If instead we required sameness
up to and including / we would make not only the past but also the present
necessary. For then only worlds exactly like the actual world at / would be acces-
sible from the actual world at / so that anything that happens at , happens neces-
sarily at l. This conflicts with the intuition that our present choices are not
determined. It must be admitted, however, that either definition of .n is metaphys-
ically presumptuous. It is a matter of whether we choose to locate the earliest point
of potential freedorn or contingency in the present moment or the immediate fu-
ture. Our semantics places it in the present.
It is not difficult to see from our definition of the accessibility relation that -e
is reflexive, symmetric, and transitive-in the sense that, for any worlds wr Lt)2,
and w, and any time / in them:
wFRwi
if wr.Rwrt, then wr.t?wrt
if wr.Rwrt and wr-Rwrt, then wr.Rwrt.
Together with the valuation rules to be stated below, this implies that the alethic
logic associated with our version of modal tense logic is 55. However, it is not this
alethic logic that interests us here, but the logic of the tense operators. The seman-
tics for this logic is given by our previous definitions together with the following
valuation rules. Rules 1-12 are just the Kripkean rules for alethic modal logic,
relativized to times as well as worlds. The novelty lies in rules 13-16. We omit
falsity rules here, since the list is getting long. But no matter; apart from cases of
nondenoting names, which we shall ignore, a formula is still false iff it is not true.
ll.ソ (□ Φ,′ ,″ )=T if fOr aH worlds z such that″ 乙κ 'zち γ(Φ ,ち 夕)=T
12.ソ (◇ Φ,ち 〃)=T ifffor some wOrld%such that″ こメ ンちソ(Φ ,ち Z)=T
l in″
13。 γ(HΦ ,ち ″)=T iff br alltimes′ ,7(Φ ,デ ,″ )=T
:ど
such that′ ″
14.ソ (PΦ ,ち ″)=T iff brsOmetime″ 'in 4′ '多 and γ(Φ ,′ ',″ )=T
15。 ソ(GΦ ,ち ″)=T if br an timesデ in″ such that″ が ,γ (Φ ,′ ',″ )=T
:in
16。 γ(FΦ ,ち ″)=T ifffor some time″ ιク;″ ″'and γ(Φ ,′ ',″ )=T
Since truth is nOw rclativized nOt Only tO wOrlds,but tO tiincs,we must once
again revise the dennitions of our chief senlantic concepts:
DEFINITION A formula is valid iff it is true at all times in all worlds on all
of its valuations.
DEFINITION A formula is consistent iff it is true at at least one time in at
least one world on at least one valuation.
DEFINITION A formula is inconsistent iff it is not rrue ar any time in any
world on any of its valuations.
DEFINITION A formula is contingent iff there is a valuation on which it is
true at some time in some world and a valuation on which it is not true at
some time in some world.
DEFINITION A ser of formulas is consistent iff there is at least one valua-
tion containing a world in which there is a time at which all the formulas in
the set are true.
DEFINITION A sel of formulas is inconsistent iff there is no valuation con-
taining a world in which there is a time at which all the formulas in the set
are true.
374 CHaprun 13
DEFINITION Two formulas are equivalent iff they have the same truth
value at every time in every world on every valuation of both.
R I_ HFR
HFR I_ HtrFR
Analysis of these inferences provides a good illustration of the uses of our seman-
tics. The first is valid. Intuitively, we can see this as follows. Suppose for reductio
that this inference is invalid-that you are now at this moment, r, reading this
book, but that it is not the case that you have always been going to read this book.
Since you have not always been going to read this book, there was a time earlier
than t, call it r', at which it is not the case that you were going to read this book.
Drorurrc arvo Terusr Locrcs 375
But then it was true at t' that at all later moments you would not read this book.
Since in particular the current moment , is later than /', it follows that you are not
reading this book now at /-which contradicts what we said above. Hence the
first inference is valid. Here is the same reasoning in strict metatheoretic terms:
卜HFR'is valid.
that this sequent is invalid;
contalnlnS sorne wor'rd w ln which there is a
"l/'
that v( OR t, w) T and y('HFR t, wl * T. Since
"lr (
.HFR" t, w) *Trby valuation rule 13 there IS SOFne tinle t'
ln w such that. t'st and * T. then by rule 76
for all tinles t" in ll) t", z) *TNow
since f is in z and ,' that Y('R', /,
w) *T, which that Y('R', /,
wl =T.
Thereお re`R卜 HFR'is valid.
' Technically, we could get by even without the past moments, but the resulting
model is so counterintuitive that it does not make a good illustration.
376 (こ HAPTER 1 3
/r. These stipulations define the model. Having done that, the only work that
remains is to apply the truth clauses for the operators to verify that this model
does indeed make the premise 'HFR'true and the conclusion'HIFR' false:
卜H□FR'is invahd
Y whose set ?of times is
tr>l and whose set of
both worlds is {tr, tr}; that is, both
are ln both worlds, and further:
v( rRt tb rur F
1r(
.R tr, w1 T
Since we are dealing here
stipulate domains for the
N ow time t2 ln world T. Hence by
rule 7 6, v( .FR tb w that
tt€tz. Hence for all tlllles t'
by rule 1 3, "l,"('HFR tb w.
Further, t2 IS the only time world w2
* T. Hence by rule 6, v .FR' uz)
(since r, IS a tlme rn both w1
the same times up to , and
up to 4. Since, then,
1 v( .nFR t1t tul * T.
by rule 13 that "l/(
that "[('FIFR', t.r, wr) =l
'is invalid. QED
Have we, then, refuted determinism? Of course not. We have refuted only
one argument for it. And this refutation employs a semantics based on controver-
sial assumptions about time and possibility. Indeed, it is a semantics designed from
the outset to represent nondeterministic time-time with many possible futures.
That it yields counterexamples to deterministic arguments is therefore no wonder.
The determinist could well retort that we have used the wrong semantics and hence
the wrong logic, that the true semantics represents time not as a branching tree but
as a single nonbranching line, and that our semantics simply begs the question.
Logic alone cannot settle this issue. For any purported solution, one can
always ask whether the correct semantics, and hence the correct logic, has been
used. But here opinions will differ. One can be extremely conservative, allowing
into one's semantics only the most strictly logical presuppositions or one can be
more venturesome, adopting presuppositions with a metaphysical tinge. (I have
been somewhat venturesome, for example, in assuming that t is transitive and,
within worlds, linear, and also in assuming that accessibility amounts to shared
Drorurrc aruo Trrusr Locrcs 377
history; the determinist, who makes the strong assumption that only one world is
possible at any given time, is more venturesome still.) conservative logics, which
operate with fewer presuppositions, are less controversial. But, because they vali-
date fewer inferences, they are also less interesting. The most interesting tense
logics venture some way into the hazy borderland between logic and metaphysics.
Though it leaves the problem of determinism unsolved, logical analysis of
arguments for determinism is not merely an empty exercise. Surely we clarify our
conceptions if we take the time to formalize them and relate them to various
models. Moreover, the fact that we can model nondeterministic time shows that
there is no contradiction in its conception-so far, at least, as we have conceived
it. Even if real time is wholly unlike our models, the existence of these models
shows that a nondeterministic universe cannot be ruled out on logical grounds
alone. There is a logic of nondeterministic time, whether or nor time itself is
deterministic.
'we
now turn to some general features of our modal tense logic. Note, first,
that tense logic reflects the most fundamental patterns of modal reasoning. In
addition to the dualities already mentioned, we have, corresponding to the rule of
necessitation in modal logic, this rule: If @ is valid then so are H@ and G@.
Similarlg corresponding to the Kripke rule K, every sequent of either of the follow-
ing forms is valid:
H(@-Y) F(HO-HY)
G(O*Y)F(GO*GY).
However, since the relation g need not be reflexive, we do not have the principles
H(DFO
GOFO
which correspond to the T rule tr@ F (D. These principles are incorrect for tense
logic since the first says that whatever has always been true in the past is still true
and the second that whatever will always be true in the future is tiue already. To
both of these theses there are easy counterexamples.
Two interesting principles that are not valid on our semantics are
P@* PP(D
F<D * FF@
The first asserts that if there was a time at which @ was the case then there was a
(presumably more recent) time at which it had already been that @; the second
asserts that if there will be a time at which @ is the case then there will be a
(presumably still earlier) rime ar which <D is going to be the case. These schemata
express the thesis that time is dense-that between any two times there exists a
third. Any model some of whose worlds contain only finitely many rimes falsifies
this thesis. since time as we experience it certainly seems to be dense, we might
want to build density inro our definition of a model. we could do this by requiring
that t be a dense as well as transitive relation-that is, by requiring that ftr any
times /, and trin a world w, if trttr, then there exists a time /, in ar such that tft3
378 Cuaprrn 13
andtrttr. That would imply, among other things, that all worlds contain infinitely
many times, and it would make the two density principles valid.
All instances of the converses of the density principles
PP<D* P@
FF@
- F<D
are, by contrast, valid on our semantics. This can be shown from our stipulation
in the definition of a valuation that, is transitive.
Tense logic displays some interesting oddities if time has a first or last mo-
ment. If there is a first moment-that is, a moment f at which time begins-in
some world w, for example, then at that world and time P@ is false for any formula
@, even if @ is valid. This is because there is no time t' in w such that t'st andT (@,
t', w) =T so that by valuation rule 14, y(PO, t, u) * T. But the real oddity is that
HO is true at t for any formula (D, even if O is a contradiction. This is because,
since there are no times f in ea such that t'rt and 1/'(<D, t', w) +T (since there are
no times t' inw such that t'ttl,itfollows trivially that for all times /'such thatt'tt,
Y(@, r', w) =T so that by valuation rule 13, "l/(HO, t, w) =1.
As a matter of fact, the formula 'PH(A & -A)" which says that there was a
time at which the contradiction 'A & -A' had always been the case, is a way of
asserting that there was a first moment in time. Similarly, 'FG(A 6{ -A)' says in
effect that there will be a last moment in time. It follows, then, that '-PH(A 6a
-A)' and '-FG(A assert, respectively, that time has no first moment and no
& -A)'
last moment. However, this is not the same as asserting that time has gone on or
will go on forever. Suppose that time will end exactly one second from now, but
that ihere still will be a moment of time % second from now, second from now,
3/+
7a second from now, and so on. Then there will be no last moment, even though
time does not go on forever. There is no way in our version of tense logic to say
that time goes on forever, since our logic deals only with the topological properties
of time (i.e., roughly, with time's "shape"), rather than with its duration. Logics
dealing with duration are called chronological logics or metric temporal logics,
but they are beyond the scope of this book.'o
Some of the valid formulas of tense logic have no analogs in the other modal
logics we have studied. Among these are the following "mixing axioms" (so called
because they mix past and future operators):
<D GP(D
-
<D-- HF(D
The first of these asserts that if <D then it will always be that O was the case; the
second asserts that if (D then it always had been that (D would be the case.
Likewise peculiar to tense logic are these axioms of past and future linearity:
(Po E PY)* (P(o & Y)v ((P(o & PY)v P(P<D & Y)))
(Fo & FY)- (F(o s( Y) v ((F(o 6. FY)v F(Fo & Y)))
)0 For an account of metric temporal logics, see Rescher & Urquhart, Temporal
Logic, chap.X.
Deourrc ano Trrusr Locrcs 379
The first asserts that if O and Y are both the case in times past, then they are both
the case at the same past moment, or Y is the case before (D, or @ is the case before
Y. The second is an analogous assertion about the future. The validity of these
principles follows from our stipulation in the definition of a valuation that time
within a world is linear.
lWe have designed our semantics (primarily through the definition of the
relation =c) to model the notion that the past is necessary. Thus we might expect
that all formulas of the form P@ - lP@ turn out to be valid. This is almost, but
not quite, true. All formulas of this form in which (D contains no occurrence of
either of the future operators 'F' or 'G' are valid. But the use of 'F' or 'G' would
allow us to make statements about what is going to be true in the future, and since
the future need not be determined on our semantics, their inclusion may result in
false instances.
To see why, let's consider a world w and a time / in ea at which Alice, standing
on the edge of a precipice with a hang glider, decides to leap-a thing she has
never done before. Suppose also that this decision was voluntary; that is, though
in z she decides to leap and actually takes the plunge, she need not have done so.
In other words, there is some other possible world z which has the same history as
w up to l, in which her decision at time / is not to leap. ln u, we shall suppose, she
stays put, deciding then and there that hang gliding is not for her and never at-
tempting it again. In w, then, at some time after r Alice is hang gliding, but in z she
never hang glides. Let's use the letter 'A' to stand for 'Alice is hang gliding.' Then
'PFA
- lPFlf , which is an insrance of the general form P<D -* CPO, is false in w at
t. For consider some time /' in w that is earlier than /. Then clearly 'Flt' is true at /'
in ra. Hence 'PFA' is true at t in w. However, 'trPFA' is false at t in w. This is because
'lf is false at all times in z. Hence in z there is no time before r at which 'Flt' is
true. So 'PFA' is false at t in u. Hence, since z is accessible from w at / (both worlds
having the same history up to ,), 'trPFlt' is false at t in w.It is, in other words, not
necessary at t in w that it was the case that Alice would hang glide-for at t, the
moment of her decision, a world in which she would never leap was still possible.
Though instances of P(D - DPO may not be valid when (D contains F or G, .
still, from those instances not containing F or G, all of which are valid, we can see
that our semantics incorporates a strong form of the doctrine of essentiality of
origin. Any object which now actually exists has, as a matter of necessity, the
entire past it actually did have, from the moment of its origin up to the present.
lVorlds whose history deviates from ours at some time before now are now no
longer possible relative to the actual world, though they may once have been
possible.
You probably have noticed that principles of tense logic tend to come in
pairs-one principle for the past and one for the future. This suggests a general
principle, known as the mirror image rule:
If valid formula, then so is the formula Y obtained from (D by replac-
@ is a
ing each occurrence of G by H and each occurrence of F by P, or vice versa.
The mirror image rule justifies, for example, an inference from 'A * GPlt' to 'A -
HFlf and vice versa.
380 CHAPTER 1 3
This rule is, in fact, valid for semantic systems that make the past and future
topologically symmetrical. Our semantics is not one of these, however, since we
have designed it to make past but not future events necessary. Consequently, not
all instances of the mirror image rule are valid on our semantics. For example,
though the formula 'PA - lPlf is valid, as noted previously, the corresponding
formula 'FA * nFlt' is contingent.
Finally, let us take note of an interesting principle involving the interaction
of quantification with modal and tense operators-Murphy's law: "Whatever can
go wrong will." Using ''W' for 'goes wrong', we might at first blush just literally
transcribe this principle as 'Vr( O !7r - F$Vx)', but that misinterprets the tenses.
It says, "'W'hatever can now go wrong will at some future time go wrong." But it is
truer to the intent to say that whatever can at some future time go wrong will at
some future time go wrong, that is:
Vr(OF\7r - FWr)
Our nondeterministic semantics provides countermodels to this thesis, but a deter-
ministic semantics does not.
Exercise 13.2.1
PrOve that the f0110、 ving sequents are valid On Our senlantics fOr n10daltense 10gic:
1.FHA卜 A
2.FHA卜 HA
3.PA卜 GPA
4.PPA卜 PA
5.HA3(PB卜 P(A3(B)
6.GA→ FA卜 ∼G(A&∼ A)
7.GHA,FA卜 HGA
H′ ″′ Sノ Zρ ογ
:I′ ′ ″″ι′
οι力θρ′οοtt ο′ι
力θルS′ ι
力″ιιργοι′
θttS ι
あαιι
あιSθ ″ι
θπιθJθ ′
″θγ
`A'ご οθs ηοιιοπι
αノ″″
あθορι″
αιο `F'o/`G':
8.PAト ロPA “
9。 ◇FHA卜 A
10。 A卜 ◇HFA
Exercise 13.2.2
Prove that the following sequents are invalid:
1,PA卜 A
2.PA卜 ◇A
3.□A卜 FA
4.□ PA卜 P□ A
5。 PA卜 PPA
Drorurrc aruo Trrusr Loctcs 381
Exercise 13.2.3
Prove that the following formulas are not valid:
Al is a frog.
Beth is a frog.
Al and Beth have something in common.
'We
might symbolize this in predicate logic by'Fa, Fb llxHabx', where'H'is a
three-place predicate meaning "- and have in common." But this
sequent is invalid and so does not capture the relevant features of the argument.
- -
What Al and Beth have in common, of course, is the property expressed by the
predicate 'F', the property of being a frog. This is the "something" referred to in
the conclusion. 'Something' normally indicates the presence of existential quanti-
fication, but in this case the quantification seems to generalize not the places oc-
cupied by the names, but the place occupied by the predicate. The form, in other
words, appears to be this:
Fa, Fb l- 3X(Xa & Xb)
where 'X' is not a specific predicate, but a variable replaceable by predicates-a
variable that stands for properties. Thus the conclusion asserts that there is a
382
HrcrEn-Onorn Loctcs 383
property X which both Al and Beth have-or, more colloquially, that they have
something in common.
If we allow such quantification, then many ideas not previously expressible
in any direct way become formally expressible. We may formalize 'Everything has
some proper ty' , for example, as 'Y xlYYx'. Notice that we now need two styles of
variables, one for properties (uppercase variables) and one for individuals (lower-
case variables). A logic which quantifies over both individuals and properties of
individuals is called a second-order logic, as opposed to systems such as classical
predicate logic, which quantify only over individuals and are therefore called first-
order logics. It is also possible to quantify over properties of properties, as in this
inference:
Socrates is snub-nosed.
Being snub-nosed is an undesirable property.
Socrates has a property that has a property.
1 This difference is usually marked by special subscripts or superscripts, but for the
simple illustrations given here, variation in size is more graphic'
384 Cuaprrn 14
Vx(0x *
-1yXy)
VX(1X *lxVy(Xy*y=x))
VX(3X *1x1y1z(((-x =y & -x = z) U -y=e) & Yw(Xw * ((w = x v w =
y)vw=e))))
and so on.
Higher-order logic enables us to quantify over these numbers and introduce
still higher-order predicates that apply to them (such as 'is prime' or 'is greater
than'). Thus it provides the resources to construct arithmetic-and much of the
rest of mathematics-in what appear to be purely logical terms.2 For this reason
Frege and also Alfred North r7hitehead and Bertrand Russell,3 writing near the
beginning of the twentieth century, held that mathemarics itself is norhing more
than logic. More precisely, they argued that all mathematical ideas can be defined
2 The details of this construction are beyond the scope of this book.
The classic
introduction to them is Bertrand Russell's Introduction to Mathematical philoso-
phy (London: Allen and Unwin, 1919).
3 The arguments of u7hitehead and Russell were
developed in their historic three-
volume work, Principia Mathematica (Cambridge: Cambridge University press,
191,0-t9t3l.
Hrcurn-Onorn Locrcs 385
in terms of purely logical ideas and that all mathematical truths are logical truths.
This thesis, known as logicism, has, however, met with serious technicaldifficulties
and is now in disrepute.a
In addition to providing an analysis of number, second-order logic provides
an analysis of identity. Identity can be defined in terms of second-order quantifi-
cation and the biconditional by a principle known as Leibniz's law. (However, we
shall not adopt this definition but will retain the identity sign as a separate, primi-
tive operator.) This is the principle that objects are identical to one another if and
only if they have exactly the same properties. In formal terms:
a=b*VX(Xa*Xb)
Leibniz's law itself is sometimes further analyzed into two subsidiary principles,
the identity of indiscernibles:
VX(Xa-Xb)-a=b
and the indiscernibility of identicals:
a=b*YX(Xa*Xb)
The first of these formulas says that objects that have exactly the same properties
are identical, and the second says that identical objects have exactly the same
properties. Leibniz's law is equivalent to their conjunction.
Higher-order quantifiers may range not only over properties but over rela-
tions as well. For example, the argument
Al loves Beth.
Al has some relation to Beth.
may be formalized as 'Lab l lZZab'.s
Analogies, to take another example, are often expressed in the form
a standsto b as c standsto d.
For instance:
a For good accounts of the demise of logicism, see Rudolf Carnap's "The Logicist
Foundations of Mathematics" and Kurt G<idel's "Russell's Mathematical Logic,"
both in D. F. Pears, ed., Bertand Russell: A Collection of Critical Essays (New
York: Doubleday,7972). See also Paul Benacerraf and Hilary Putnam, eds., Phi-
losopby of Matbematics: Selected Readings,2nd ed. (New York: Cambridge
University Press, 1 983).
5 It is common practice in higher-order logic to mark variables standing for n-place
predicates with the superscript or subscript 'z'; but since we require that a quanti-
fied variable always occurs both with the quantifier and at least once later in the
formula, this is unnecessary, for we can tell which sort of predicate the variable
stands for by counting the names after one of the later occurrences of the variable
in the formula. Thus we know that in')ZZab','Z'is avariable ranging over two-
place relations, for in its second occurrence it is followed by two names.
386 CHAPTER 1 4
the analogy here being the relationship between a country and its capital city. The
general assertion that there is such an analogy among four particular items may
be formalized as
lZ(Zab k Zcd)
This is just like the quantifier clause for predicate logic, except that it concerns
predicates and predicate variables, rather than names and individual variables.
Exercise 14.1
Formalize the following arguments in second-order logic, using the interpretation
suggested below. All of these arguments are valid.
HrcHrn-ORorR Locrcs 387
1. All humans are mortal. Therefore all humans have something in common.
2. Al is human. No humans are perfect. Therefore there is at least one property
which Al has and at least one which he doesn't have.
3. All objects are such that there is no property which they both have and lack.
Therefore, for any object and any property, either it has that property or it does
not.
4. Al is Bud. Therefore Al and Bud have exactly the same properties.
5. Everything is either human or nonhuman; therefore there is a single property
which every object has.
6. Everything is identical to itself; therefore there is some relation which every-
thing has to itself.
7. Al loves himself. Therefore not all relations are irreflexive.
8. Some relations are asymmetric. Therefore some relations are irreflexive.
DEFINITION Let V be any valuation, Y any n-place predicate (z > 0), and
4^t any set of a-tuples of objects in the domain of V. Then Y potentially
characterizes -v with respect to V if and only if V does not assign some
extension other than .,tr to Y . (In other words, Y potentially characterizes tv
if and only if either'f (Y) = .tv or Y assigns nothing to Y.)
DEFINITION Let V be any valuation, .,\/ arry set of z-tuples (z > 0) from the
domain of 1/, and Y any n-place predicate which potentially characterizes -w
with respect to V. Then the second-order expansion Y,*.=, of Y with respect
to -w and Y is the valuation which has the same domain as .I/ and which
assigns the same extensions to the same predicates and names, but which lz
addition assigns to Y the extension t, if Y does not already assign Y an
extension. If Y already assigns .Nto Y )then Y1v,-e is just "l/.
Now to the valuation rules of first-order predicate logic, we add the follow-
ing new clauses. Let A be any n-place predicate variable (n > 0) and @(Y/A) the
result of replacing all occurrences of the A in <D by Y. Then:
11. l/(VAO) = T iff for all sets -u of n-tuples from a and for all n-place
predicates Y that potentially characerize =,v, V,*.-.,(@(Y/A)) = T;
Y(VAO) = F iff for some set -4lof a-tuples from r and some n-place
predicate Y that potentially characterizes -lr, y(,y,_e(O(Y/A)) + T.
12. y(lAO) =T iff for some set =nrof z-tuples from n and some n-place
predicate Y that potentially characterizes -^/,'f(.",_e(O(Y/A)) = T;
HrcurR-Onorn Locrcs 389
Y(IAO) = F iff for all sets -u of n-tuples from a and for all n-place
predicates Y that potentially characterize -/v, "l/(v,=^r(O(Y/A)) * T.
Using this semantics, we can see, for example, that the formula 'lXXa',which we
considered earlier, is not merely true on some particular valuations but is a valid
formula.
`ヨ XXa'is valid.
Trees in second-order logic are just like those for first-order predicate logic,
except that they employ four new rules-two each for each of the two new second-
order quantifiers. Except for the universal quantifier rule, they are little different
from the corresponding rules for first-order quantifiers. The existential rule is:
Second-Order Existential Quantifier (23) If an unchecked formula of the
form 3AO, where A is an n-place predicate variable, appears on an open
path, check it. Then choose an n-place predicate Y that does not yet appear
anywhere on that path and write @(Y/A), the result of replacing every oc-
currence of A in @ by Y, at the bottom of every open path that contains the
newly checked formula.
'We
may use this rule to show, for example, that the form'3X(Xa & Xb) l- a = b'is
invalid. (Just having at least one property in common does not guarantee identity!)
Here is the tree:
1./∃ X(Xa&Xb) Premise
2. ∼a=b Negation of conclusion
3./Fa&Fb t21
4. Fa 3&
5。 Fb 36.
The tree exhibits a counterexample in a domain of two distinct individuals, both
of which have the property F.
The tree rules for negated quantifiers are exactly like those for first-order
predicate logic, except that they mention predicate variables instead of individual
variables:
390 CHaprrn 14
The following tree uses 2-Y and2l to show that'a =b f VX(Xa - Xb)' is valid
1.a=b Premise
2. /∼ ∀X(Xa→ Xb) Negation of conclusion
3./∃ X∼ (Xa‐→Xb) 22-V
4. /∼ (Fa― →Fb) 321
5。 Fa !-+
6。 ∼Fb !--
7. Fb 1,5=
8. × 6,7
The second-order universal quantification rule, however, is more compli-
cated. To see the need for the complication, consider the following example. It is
certainly true (indeed valid) that if Al has any property then Al does not nor have
that property. In symbols: vX(Xa
- --Xa).It ought to follow in particular that if
Al is loved then Al is not not loved. But the property of being loved is, on at least
one useful formalization denoted not by a simple predicate, but by the complex
expression '1xLxy', where 'Ij is a two-place predicate meaning "loves." The com-
plex expression readsr "There is something that loves yi'or, more simply, "y is
loved." Correspondingly,'Al is loved' is'].xLxa'. Thus what follows as an instance
of 'VX(Xa - --Xa| is 'lxlxa
- --1xLxa'.
Notice that to obtain this instance we had to replace the predicate variable
'X' in 'VX(Xa - --Xa)' not by a single one-place predicate, but by the entire
expression '1xLxy'.In the process, the place of the unquantified variable 'y' was
taken by the name 'a'.
Now let's consider a more sophisticated example. It is also true (and valid)
that for any relation, if Al has that relation to Beth, then something has that
relation to Beth-that is,YZ(Zab
- 1xZxb). And from this it ought to follow
(since nonidentity is a relation) that if Al is not identical to Beth, then something
is not identical to her-that is, -a = b * lx-x = b. Now the relation of nonidentity
is expressed by the complex formula '-x = !'. To obtain '-a = b
instance of 'Y Z(Zab 1xzxb)'
-1x-x = b' as an
- , we remove the quantifier and replace each occur-
rence of the predicate variable 'Z' by '-x = y', substituting the first name or vari-
able following that occurrence of 'Z' for 'x' and the second name or variable
Hrcrrn-Onorn Locrcs 391
following that occurrence of 'Z' for'y'. It is this sort of substitution that makes
the universal quantification rule somewhat difficult to express.
Notice that expressions such as 'lxLxy' and '-x = y' are not officially for-
mulas, since they contain unquantified variables. They are, rather, expressions that
could be obtained by removing initial first-order quantifiers from quantified for-
mulas. The unquantified variables in such expressions are known as free variables,
and such an expression containing exactly n free variables is called an n-place
open sentence. Thus '1xLxy'is a one-place open sentence, because it has one free
'What
variable, 'y' , and '-x = !' is a two-place open sentence. we count in deter-
mining the number of places is the number of distinct variables, not the number of
occurrences of variables. So, for example, 'Lxx v -Lxx' is a one-place open sen-
tence, even though its one variable occurs four times.
An instance of a second-order universally quantified formula, then, is a for-
mula obtained by removing the quantifier with its attached n-place predicate var-
iable and replacing remaining occurrences of that variable with an n-place open
sentence so that names and variables wind up in the right places in the manner
illustrated above. More precisely:
Being an instance is an easier thing to show than to say! Having gotten this far,
however, we should not find the universal quantifier rule too difficult:
/
∃喜
・
Premise
ヽ
v一
2
Xa
´
Negation of conclusion
3
L111
a
”×
L-)
.
4
32Y
5
1,4
392 Cuapren 14
Here we may think of the one-place open sentence that replaces the predicate
variable 'X' at step 4 simply 2s 'p;'-s1 'Fy', and so on: The choice of the lower-
case variable is arbitrary. Notice that in accordance with the rule 'Fx'contains no
names or predicates not already on the path.
Sometimes, however, the path may contain no names or predicates at all. In
that case we can still form instances by using the identity predicate, for which
explicit exception is made in the rule. Consider this tree, which shows that the
assertion that something has some property,'1X1yXy', is valid:
1. /∼ ∃Xヨ γXy Negation of given
2. ∀X∼ ∃γXン 1 2-=
3. /∼ ヨン
ツ=γ 22V
4. /∀ y∼ y=ソ 3-l
5. ∼a=a 4V
6. × 5
Here we may think of the open formula replacing 'X' at step 3 as'x = x' .
Finallg let's use the tree rules to demonstrate a mildly surprising result-
namel5 that being identical to something amounts just to having every property
that thing has (or, more properlS belonging to every set it belongs to). That is, we
are going to show that the formula 'a = b * VX(Xa -- Xb)' is valid. This is an
apparently stronger version of Leibniz's law, which asserts that two things are
identical iff they have exactly the same properties, that is: a = b * VX(Xa * Xb).
It turns out, however, that if b has all of a's properties, then as a matter of logic a
has all of b's-so that the two formulations are really equivalent. Intuitively, we
can see this as follows. Suppose for reductio that b has all of a's properties but that
a does not have all of b's. Then there is some property F that b has but a doesn't.
Since a doesn't have F, then a has the properry -F. And since b has all of a's
properties, it follows that b has -F, and hence has both F and -F-a contradiction.
That explains the equivalence of 'YX(Xa
- Xb)' and 'VX(Xa * Xb)'. Here is the
tree that shows that 'a = b' and 'VX(Xa -* Xb)' are equivalent (i.e., that'a = b *
VX(Xa .- Xb)' is valid):
7. / -(a = b * VX(Xa - Xb)) Negation of given
2.a=b 1∼ →
∀X(Xa・→ Xb) 1∼ →
The trick here is in the right branch at step 4 where we instantiate the predicate
variable 'X' in 'VX(Xa .- Xb)' with the one-place open sentenc e 'a = l' to obtain
'(a= a + a = b)'. 'VX(Xa - Xb)'asserts that b has every property that a has. We
take that quite strictly, concluding that if a has the property being identical with a
(expressed by the formula'a = y'), then b also has this property.
Second-order logic has a big surprise in store when we consider the questions
of soundness, completeness, and decidability. Second-order logic is undecidable,
of course, since it contains first-order predicate logic, and we proved in Section
10.6 that first-order predicate logic is undecidable. That means that some invalid
forms in second-order logic have infinite trees. No surprise there. But though the
tree rules for second-order logic are provably sound, it can be shown that no
validity test for second-order logic is complete. That is, every test, including the
tree test, fails to recognize some valid forms as valid. This surprising and frustrat-
ing fact is a consequence of Godel's historic proof of the incompleteness of arith-
metic, which is beyond the scope of this book.7
Exercise 14.2.1
Prove the following metatheorems
l. 'YXXa l- Fa'is valid.
2. 'Fa F VXXa'is invalid.
3. 'i-XYyXy'is valid.
4. 'lXYy-Xy'is valid.
Exercise 14.2.2
Test the following formulas for validity, using trees:
' If you would like to pursue these matters, I heartily recommend George Boolos
and Richard Jeffrey, Computability and Logic,2nd ed. (Cambridge: Cambridge
University Press, 1 980).
││ ││. ‐
‐
ρハ
.
・
RT
│ │■ .
V
‐
││ │ ││ │ ‐ ‐ ‐
‐
││ │ ││ │‐ ‐
‐
●︱ ●︱ 一● 一一 ●
■ ││ ‐
││ │‐ .│ .
・ ‐ .■ ‐
││ ││
‐ _ ‐ ‐‐
‐
││ │││ ││ .‐ .
‐
■│■ │ ││ ││
‐ │
│
‐ │
■ .│ .. . .‐
││
..
・ │ ‐
│‐
_
‐
‐ ‐│‐ _ . .
‐
‐
‐ ‐ .│ ││ │││ ││‐
│.││
‐
‐ │ │ ││ ││‐
││ ││
‐‐ .‐ ‐ ‐ ‐
‐‐ │ │ ■│.
CHハ PTfR ■5
MILDEY NONCLASSICAL
LOGICS
All the logics we have examined until now have presupposed that (1) the
logical meaning of a statement is its truth conditions, (2) all statements
are either true or false but not both, and (3) every name refers to some
existing thing. The logics we have developed under these assumptions
are called classical logics. But while these assumptions are justifiable for
some applications of logic, for others they are not. ln this chapter and
the next we raise specific challenges to these assumptions. The result,
as we shall see, is not anarchy or chaos but an intriguing plurality of
nonclassical logics. ln this chapter we focus on some relatively tame
departures f rom classicism (hence the chapter title). ln the next we shall
consider more radical deviations. This division is, however, somewhat
arbitrary, and some logicians regard some of the developments even of
this chapter as beyond the pale.
"something exists," is also classically valid. Its validity results from the stipulation
in the classical definrtion of a valuation that domains are nonempty. This stipula-
tion is made to ensure that names have something to denote. But, while true,
'1x x = x'seems nor ro be a genuine logical truth. "'Why is there something, and
not rather nothing?" asks the philosopher Martin Heidegger; and the answer, if
there is one, surely does not lie in the semantic conventions of classical predicate
logic. No merely logical response can placate such questioning. A more modest
Iogic would allow empty domains and would therefore have no such theorems as
'1x x = x'.
There is good reason, then, to seek a logic capable of handling nondenoting
names, names denoting objects that do not exist at a given world, and, perhaps,
empty domains. Such a logic is called a free logic, because it is free from the
presupposition that a name must denote some existing thing. The development of
free logics has followed two different strategies.
The first strategy is simply to allow some names to have no denotation at all.
This seems most appropriate in the case of fictional names or theoretical names
(like 'the ether') that turn out to have been based on misconceptions. This strategy
requires a modification of the valuation rules for atomic sentences containing
names, since it is ordinarily presupposed that the names denote something. In
classical predicate logic, for example, we expressed the valuation rule for atomic
formulas with one-place predicates as follows:
1. If @ is a one-place predicate and cr is a name, then V(<Dc) = T iff Y(cr) e
v(o).
But where cr has no denotation, Y(c) does not exist, so this rule must be supple-
mented by some additional convention.
There are various ways of handling nondenoting names. !7e might, for ex-
ample, make Y(<Dct) false whenever ct does not denote. Or we might allow either
truth value to be assigned to such formulas-modeling, for example, our intuition
that although Hamlet does not exist, it is true that he is a character in one of
Shakespeare's plays, but false that he is buried in Rome. We might also make (Dct
truth-valueless when c does not denote. This third choice is best implemented in a
supervaluational semantics of the sort discussed in Section 15.3. There it will be
seen on at least one natural rendering to lead not to free logic, but to a classical
logic with a nonstandard semantics. The first two options for handling nondenor-
ing names are problematic in various ways and will not be considered here.l
The second strategy for developing free logics is to retain the idea that all
names denote but to represent some of the things names may denote as not exist-
ing. This is accomplished by defining two domains: one representing actually ex-
isting objects, the other representing both existing and nonexisting objects. Names
may denote any object in the second and wider domain, but the familiar existential
and universal quantifiers range only over the narrower domain of existing things.
1 An excellent survey of the options is found in the essay "Free Logics" by Ermanno
Bencivenga in D. Gabbay and F. Guenthner, eds., Handbook of Philosophical
Logic, Volume lll: Alternatiues to Classical Logic lDordrechr: D. Reidel, 1985).
lVrrory Nor'rcusstcnl Loctcs 399
one-place predicate O, y(@, arz) is now a subset of ry rather than of a*,. Similarly,
each name denotes an object in 7 though the extension of a name-in conrrasr to
that of a predicate-is always rigid, that is, nor world-relative. On Meinongian
semantics it becomes part of the task of constructing a model to stipulate for each
predicate and each world not only which existing objects that predicate is true of
in that world, but also which nonexisting objects it is true of in that world. We
define extensions of predicates at worlds not simply for all the actual objects in
those worlds, but for all nameable objects.
Such a semantics does not automatically decide whether it is true or false
that Arkansas is a state in a world in which Arkansas doesn't exist. 'We must
make that decision in defining the extension of the predicate 'is a state' for the
various worlds. The choice itself seems largely a matter of convention or stipula-
tion. But the semantics does allow us to model this choice once we have made it.
There is nothing in the classical semantics of Section 11.2 that allows us even to
model it.
While the extensions of predicates and names are chosen from the outer
domain 7 we retain the standard quantifier rules, according to which quantifiers
range only over the inner domain b-. Thus, if we decide to make 'Sa' ("Arkansas
is a state") true in some world w where neither Arkansas nor any other state exists
(i.e., no member of V('S', rz) is in 1,,), then even though'Sa'is true in z,'3xSx'is
false there. So one way in which Meinongian semantics differs from classical se-
mantics is that it makes the rule lI invalid.
Since nothing essential to the understanding of free logic hinges on whether
we take a Kripkean or Leibnizian approach to modality, we shall here adopt the
simpler Leibnizian approach, ignoring the accessibility relation -,?. 'We thus char-
acterize a valuation "[ as follows:
'We
may now complete the valuation rules for atomic formulas containing n-place
predicates, z > 0. These rules need no longer be restricted to the case in which the
extensions of the relevant names are in b.. Instead, they now read simply:
such that.││‐
γ (`a')=1 ││││‐
│││‐
‐
│.
tllc QED
Because inner domains may be empty, as the previous metatheorem illustrates, the
logic itself does not presuppose the existence of anything. In particular, the asser-
tion that bomething exists, '1x x = x' is no longer valid as it is in classical logic:
4OZ Craprrn 1 5
ウ={1}
ヽ″={}
Since D- = { }, there is no name cr of some object I in o- such
Y1",/1(e = q, w) = T. ThereforeY('1x x = x', w) *T, and so
'1x x = r' is not valid. QED
Formulas of the form 1x x = o are true at a world w iff Y(a) € r:_-that is,
iff the object denoted by o is a member of the set of things that actually exist at u.
Thus lx x = ct means "ct exists." It is customary to abbreviate such formulas using
the symbolism 'E!ct', which, again, may simply be read as "ct exists." We saw with
our first metatheorem that existential introduction (or existential generalization)
is invalid on a Meinongian semantics, but a restricted version of it still holds, as
the following metatheorem shows:
(again by of an expansion)
So Y(ct) = 1. Hence cr. is a
I
‐.
ThuS'SinCe Φα C° ntainS αandツ (Φ α
/β /8,″ )=■ 7(コ χ Φ,η )│■ │
li COntrary to Our supposition. .‐ ‐ ‐
│ ││││ ・ │
Therefore all sequents of the form (D*/u, Ela l- lg<D are valid. QED
So, for example, we might reason concerning a particular rabbit named Allison:
Allison is a rabbit.
Allison exists.
Some rabbit exisrs.
lVrrorv Noncrnssrcar Locrcs 403
In symbols:
Ra, Ela F lxRx
This is a simple sequent of the form <D"/u, Elcr F lB(D, which is the subiect of the
preceding metatheorem.
Having provided a Meinongian semantics for quantified Leibnizian modal
logic, we can now take care of the unfinished business of supplying inference rules
for the quantifiers of Leibnizian modal logic. The rules given here apply to free
logic in general, not ,ust to modalized forms of free logic. As with classical predi-
cate logic, we need an introduction and elimination rule for each of the quantifiers.
But all four of these rules require slight modification. Since, for instance, the name
'a' need not denote anything in the actual or present situation, to infer'lrFx'from
the premise 'Fa', we need the additional premise 'lx x = a'. Similarly, to infer 'Fa'
from the premise 'V.rcFx', we need that same additional premise. Indeed, each of
the classical quantifier rules must be modified by the addition of this kind of
existence statement at one place or another:
By adding FfI, FlE, FVI, and FVE to the seven rules listed in Section 11.4,we
obtain a logic that is sound and complete with respect to the Meinongian seman-
tics outlined here. That is, a sequent of modal predicate logic is provable in that
logic iff it has no counterexample in our Meinongian semantics. Like classical
4O4 Crapren 1 5
predicate logic, however, and for essentially the same reasons, this logic is
undecidable.
'We
now illustrate this inferential system with a few simple proofs. We begin
with the sequent'Vr(Fr - Gx), Fa,1x x = allxGx':
L
↓
G
∀ F
′
1
A
χ a
χ
2
A
3
ヨχχ=a A
4
Ga 2,4-→ E
6
∃χGχ 3,5F∃ I
\Without the premise '1x x = a', this proof would fail, for 'a' might refer to a
nonexisting object that has the property F but not the property G, even though all
actual things having F also have G. Notice that, unlike their classical counterparts
which utilize only one premise each, FVE and FII always require two premises.
Uses of these rules must therefore be annotated with two line numbers.
This proof of the theorem 'F Vr(Fx * fyFy)' illustrates the use of the rule
FVI:
1
1xx=a H(brF∀ I)
2
lFa H(br→
|
3
:yry 1,2Fヨ I
4
2-3-→ I
Fa
- lyFy
5
Vx(Fx
- lyFy) l-4F∀ I
﹂
1
A
χ χ
2
Fa*.lxx=a H(for Fヨ E)
3
Fa 2 8ζ E
4
1xx=a 8ζ E
ー ヨ
6
fxOFx .5F∃ I
一
鋤
7
,2-6F∃ E
2 This sequent, which was proved in Section 11 .4, is used here as a derived rule in
the manner described in that same section.
N/torv NoNcussrcar Locrcs 405
VxFx H(fOr→ 1)
2
|I -F, H(br→ I)
3
I lxx=a H(br∼ I)
llFa
4
1,3F∀ E
I I
5
Fa&-Fa 2,4&;I
6
| -1**=^ 3-5∼ I
7
1. □∀χFχ A
2。 ◇∃χχ=a A
3。 □(∀ χFχ ―
→ (∼ Fa― →∼ヨχχ=a)) N(∀ χFχ ―)(∼ Fa→ ∼ヨχχ=a))
4. □∀χFχ → □(∼ Fa― →∼∃χχ=a) 3K
5。 □(∼ Fa―→∼∃χχ=a) l,4¨ →E
6。 □∼Fa― →□∼∃χχ=a 5K
7. ∼□∼∃χχ=a 2 DUAL
8. ∼□∼Fa 6,7R4T
9。 ◇Fa 8 DUAL
■Жercise 15.1.1
(〕 OnStruCt pr00fs for the following sequents in free Leibnizian predicate logic:
1. lxx=a,-Fa F -V;Fr
2. VxaFx F VxFr
3. VxFr F VxOFr
4. lYxx=x
5. F -lrO -x=x
6.1xx=all.xx=x
7, 1x-Fx F -VxFx
8.]xx-al]x.x-a
9. F (Fa & 3y a =y) -1ypy
10. 3yr(Fy k 1x x = y) l- n3yFy
406 CHAPTER 1 5
Exercise 15.1.2
Prove the following metatheorems:
In Section 3.1, I summarily announced that we would consider only two truth
values, T and F, and that we would assume that each statement had exactly one of
'We
these truth values in every possible situation. have followed that policy ever
since. But there are reasons not to be satisfied with it. There are, for example,
sentences which, though grammatically well formed, seem to have no intelligible
meaning. Suppose someone says, for example, "Development digests incredulity."
This might be a striking metaphor of some kind, but what it means, if anything, is
hardly clear. One might argue that it is simply false. Then 'Development does not
digest incredulity' is true. But this statement is hardly more intelligible than the
first. We might, then, decide to rule them both out of court; they both, we might
conclude, are neither true nor false. Against this conclusion, some have obiected
that since neither sentence is really intelligible, neither makes a statement, and
where no statement is made there is nothing that can be either true or false. The
question of truth or falsity simply does not arise. On this view, unintelligible sen-
tences do not challenge the principle of bivalence because that principle applies
only to statements.
A different, and perhaps stronger, challenge to the principle of bivalence
stems from reference failure. Consider, for example, the statement 'The Easter
Bunny is a vegetarian'.3 In one sense, this is not true, since there is no Easter Bunny.
But if we conclude that it is false, then we must assent to 'The Easter Bunny is not
a vegetarian'. Is he, then, a carnivore? The problem, of course, is that since the
term 'The Easter Bunny' does not pick out an object of which we may predicate
either vegetarianism or nonvegetarianism, it seems misleading to think of these
sentences as either true or false. \We might reasonably conclude, then, that because
of the reference failure they are neither. Notice that here it is less plausible to argue
that no statement has been made. We understand perfectly well what it means to
say that the Easter Bunny is a vegetarian. But what is asserted seems not to be
either true or false.
I We are here regarding'The Easter Bunny'as a proper name, rather than a definite
description. Thus it is by no means obvious that we could successfully deal with
this reference failure by appealing to Russell's theory of descriptions.
Nrltorv Norucrassrcar Locrcs 4O7
Sentence S is false.
Here we can actually offer a metalinguistic proof that it is neither true nor false.
For suppose for reductio that it is true. Then what it says (that it is false) is correct,
and so it is false. It is, then, on this supposition both true and false, which contra-
dicts the principle of bivalence. Suppose, on the other hand, that it is false. Then,
since it says of itself that it is false, it is true. Hence once again it is both true and
false, in contradiction to the principle of bivalence. Hence, from the principle of
bivalence itself, we derive by reductio both the conclusion that this senrence is not
true and the conclusion that it is not false. It is, then, certainly not bivalent.
One might also reject bivalence on metaphysical grounds. Jan l.ukasiewicz,
who constructed the first multivalued semantics early in the twentieth century,
held that contingent statements about the future do not become true until made
true by events. Suppose, for example, that a year from now you decide to write a
novel. Still it is not true now that a year from now you will decide to write a novel;
the most that is true now is that it is possible thar you will and possible that you
won't. Only when you actually do decide a year hence does it become true that a
year earlier you were going to decide to write a novel a year hence. ObviouslS
Lukasiewicz's conception of time is different from that presenred in Section 13.2,
where we modeled a nondeterministic time in which contingent statements about
the future may be true at present. l.ukasiewicz assumed that the present truth of
contingent statements about the future implies determininsm. In any case, the idea
that the truth of a contingent statement does not "happen" until a specific moment
in time, whether right or not, is of logical interest. It implies that many statements
about the future are neither true nor false now so that there is some third semantic
status, which Lukasiewicz called 'possible' or 'indeterminate', in addition to truth
and falsity.
There may also be more mundane, pracical grounds for rejecting bivalence.
The designers of a computer database of propositions, for example, might want to
Iist some propositions as true, some as false, and others as unknown There is, of
course, no metaphysical basis for the third value in this case. The propositions
listed as unknown may all in fact be true or false. But in practice the inferen-
tial routines used with the database may work best if they embody a non-
bivalent logic.
Finally, phenomena of vagueness might motivate us to reject bivalence. The
seemingly simple and clear sentence 'This is a car' asserted of a midsize sedan is
certainly true. Asserted of an eighteen-wheeler it is certainly false. But what if it is
asserted of a van? Many people feel that such assertions are "sort of true, but not
exactly true." Since, like 'car', virtually all words are somewhat vague) for virtually
all statements there are borderline situations in which we are hesitant to say that
the statement is either true or false. But the notion that truth comes in degrees
leads beyond consideration of a mere third alternative to truth and falsity into the
realm of infinite valued and fuzzy logics (see Section 16.1).
408 CHAPTER 1 5
The difference between these two policies is easily seen in the case of disjunction.
Suppose that the sentence letter 'P' has the value T and that the letter 'Q' has the
value I. Then what is the truth value of 'P v Q'? According to the first policS the
indeterminacy of 'Q' infects the whole formula so that the truth value of 'P t, Q'
should be L But on the second policy, we must recognize that the truth of either
disjunct is classically sufficient for the truth of the disjunction, and so the truth
value of 'P t Q' should be T.
It is not obvious which of these policies is preferable. Indeed, one may be
preferable for some applications, the other for others. Thus we shall discuss them
both and, toward the end of this section, add a third to the mix.
In the late 1930s the Russian logician D. A. Bochvar proposed a three-valued
semantics for propositional logic based on the first policy-namely that the inde-
terminacy of a part infects the whole. This semantics is encapsulated in the
following truth tables:
(D
-ooY o&Y @vY O.-Y @*Y
T
T
T
T
T
T
T
T
T
F
F
T
F
F
F
T
I
I
I
I
I
I
I
F
F
T
T
F
F F I
T
T
F
F
I
I
I
I
I
I
I
T
I
I
I
I
I
I
F
I
I
I
I
I
I
lVtrorv Nottcmssrcal Locrcs 409
Notice that, whereas in classical logic for formulas or formula sets containing z
sentence letters there are 2" valuations (horizontal lines on the truth table), in
three-valued logic there are 3". The truth table for a binary operator, for example,
has nine lines.
The valuation rules for Bochvar's logic reflect this new complexity. I7e shall
for'-' and '&'may serve as examples:
not state them all, but the rules
Y(-O)=TiffY(<D)=F.
y(-o)=Fiffv(@)=T.
1/(-O)=Iiff"l/(o)=I.
"l/(o & Y)=T iff bothv(o)=T andV(Y)=f.
1/(O 6. Y) = F iff eitherl/(O) = F andY(Y) =l ory(O) =T andYlYl = q
or 1/((D) = F and "f(Y) = F.
1/(O & Y) = I iff either V(@) = I or T(Y) = I, or both.
The resulting logic has some striking features. For one thing, all of those formulas
which were classically tautologous (i.e., valid by their truth tables) are on Boch-
var's semantics tautologies no longer. The classical tautology 'P -- P', for exam-
ple, has the value I in the case in which 'P' has the value I. It is not, therefore,
true on all valuations and hence not a tautology. In fact, every formula of ordi-
nary propositional logic is indeterminate on any Bochvar valuation in which
any one of its sentence letters is indeterminate. Those formulas which were tau-
tologies in classical logic still, however, have a special character: There is no valu-
ation on which they are false. But they are not tautologies, as we have defined
that term.
We might, of course, broaden our definition of tautology. \We could say that
a tautology is any formula which is not false on any line of its truth table-that is,
which is either true or indeterminate on all lines. Then the classical tautologies
would count as tautologies, even on Bochvar's semantics. Thus it comes down to
a question of which truth values we shall accept as designated-that is, which
values count toward tautologousness. If a statement must be true on all lines of its
truth table to count as a tautologS then T is the only designated value. If a state-
ment need merely be either true or indeterminate on all valuations, then both T
and I are designated values. For Bochvar, only T was designated.
The definition of validity itself presents us with similar choices. In a bivalent
logic, it makes no difference whether we say that a sequent is valid iff
there is no valuation on which its premises are all true and its conclusion is
untrue
or
there is no valuation on which its premises are all true and its conclusion is
false.
Given bivalence, untruth and falsity are the same thing. In a multivalued logic,
however, the difference between the two definitions is substantial, for there may
410 Cuapre n '1
5
be valuations on which the premises are true and the conclusion is indeterminate.
Should these count as counterexamples? If we think so, we will adopt the first
definition of validity. If we think not, we will adopt the second. Bochvar adopted
the first, and it is the one we shall use here. lndeed, we stipulate now that ue
sball for tbe purposes of tbis section (and particularly the exercise at tbe end)
retain the wording of all the definitions of semantic concepts presented in Cbap-
ter 3.
Another interesting feature of Bochvar's semantics is that it invalidates a
number of sequents that at least some logicians have regarded as suspect. The so-
called "paradoxes of material implication," for example:
Qrr'-.q
-PFP-.Q
though valid in classical logic, are on Bochvar's semantics invalid. In the first case,
the counterexample is the valuation on which 'Q' is true and 'P' indeterminate,
which makes the premise true and the conclusion indeterminate (hence untrue). In
the second, the counterexample is the valuation on which 'P' is false and 'Q'
indeterminate. Indeed, the invalidity of these sequents is an instance of the follow-
ing general principle:
METATHEOREIⅥ : Lct Φl Φ ト Ψ bc a
pr6positiona1 1。 gic as defined by the
Then on Bochvar.ゝ senlantics, if {@,,' Φ″
} Ψ con―
tains a sentence letter not found Φ″
, then that seq uent lS
illv‐ alid.
In stating this metatheorem I said explicitly that the formulas in question were to
be formulas "of ordinary propositional logic (as defined by the formation rules of
Chapter 2)." The reason for this qualification is that Bochvar added a novel oper-
ator to his logic, and the metatheorem does not apply to formulas containing this
novel operator.
lVuorv Noruclassrcal Locrcs 411
Following Susan Haack, I shall symbolize this operator as 'T', since it means
in effect "it is true that."a Like negation, it is a monadic operator, so to add it to
our language we adopt this new formation rule:
If <D is a formula, so is TO.
Φ TO
T
T
F
F F
I
That is, if a formula (D is true, then it is true that (D. But if O is false or indetermi-
nate, then it is not true that <D.
This "truth operator" gives Bochvar's logic a new twist, for any formula of
which it is the main operator is bivalent. And though, as we saw above, Bochvar's
logic contains no tautologies among the formulas of ordinary propositional logic,
it does have tautologies. These, however, are all formulas containing the truth
operator. Here are some examples:
TP-TP
TP v -TP
T(P v -P)v (-TP & -T-P)
Notice, by contrast, that'TP v T-P' is not tautologous, for it is false when 'P' is
indeterminate.
Bochvar hoped that his new semantics would solve the problem of semantic
paradox. Consider, for example, the semantically paradoxical sentence that we
have called S:
Sentence S is false.
I[e argued above that S is neither true nor false. But if we adopt Bochvar's seman-
tics, there is a third option: It might have the value I. Suppose, then, that it does.
In that case, using the sentence letter'S' to represent sentence S, though the for-
mula 'S'has the value I, the formula 'TS' has the value F, for it is in fact not true
that sentence S is false.
This three-valued approach seems neatly to dissolve the paradox. Unfortu-
nately, however, a new paradox emerges to take its place. Let U be the sentence:
Sentence U is untrue.
As before, suppose for reductio that this sentence is true. Then what it says is
correct, and so it is untrue. Hence it is both true and untrue-a contradiction.
Therefore it is not true. It follows, on Bochvar's semantics, that it has one of the
values F or I. But in either case it is untrue, and so what it says is correct; hence it
is true. Once again we have a contradiction-despite the third value.
Bochvar's semantics does not, therefore, provide a general solution to seman-
tic paradoxes-nor does any other three-valued or multivalued semantics. If se-
mantic paradox is the problem, multivalued semantics is not the solution.
Bochvar, as we have seen, settled upon the first of the valuational policies
mentioned above-the policy of making indeterminacy of a part infect the whole.
The second policy is to assign truth values to the whole even if some components
are indeterminate, so long as the truth or falsity of the remaining components is
sufficient to determine the truth value of the whole by the classical valuation rules.
This second policy results in a different set of truth tables:
T T T
T F I
F T I
T
T
T
T
F
F
F
T
I
I
T
T
F
F
T
F
T
F
T T I I
I T I I
I
F
F
T
I
I
I
I
I
The three-valued propositional semantics expressed by these tables was first pro-
posed by S. C. Kleene. (Kleene also investigated Bochvar's system independently.)
As with Bochvar's logic, we may also introduce the operator 'T', which has the
same truth table as before.
On Kleene's semantics, as on Bochvar's, classically tautologous formulas are
nontautologous. For, as on Bochvar's semantics, any statement of ordinary prop-
ositional logic all of whose atomic components have the truth value I has itself
the truth value I; hence, for any formula, there is always a valuation (namely,
the valuation that assigns I to all of its sentence letters) on which that formula is
not true.
Yet Kleene's logic differs from Bochvar's in that it makes most of the classical
inference patterns, including the paradoxes of material implication, valid. There
are, however, some exceptions: In particular, inferences to what are classically
tautologies from irrelevant premises still fail. The sequent 'P l- Q * Q', for exam-
ple, though classically valid, is invalid on Kleene's semantics, for the valuation on
which 'P' is true and 'Q' indeterminate is a counterexample.
Kleene's semantics assigns the classical values T and F to more complex
formulas than Bochvar's semantics does. [.ukasiewicz, who was the first to explore
three-valued logic, proposed a semantics that goes even further in this direction.
Lukasiewicz's semantics is like Kleene's, except that where Kleene makes the con-
ditional and biconditional indeterminate when both their components are indeter-
minate, Lukasiewicz makes them true. l.ukasiewicz's semantics is thus expressed
by the following truth tables:
lt/rrolv Noncrnssrcar Locrcs 413
T
T
T
F
T
T
T
T
F
T
F
F
F
T
I
I
I
I
I
T
T
T
T
F
F
T T T
F I T
F I T
T
F
I
F
I
I
I
I
F
I
I
T
I
I
T
I
formulas into those that are meaningful and those that are meaningless, assigning
the latter the value I. Or, if we are doing tense logic, we might design our models
so that atomic statements about the future always receive the value I. But again,
we shall not bother with the details of these variations.
Exercise 15.2.1
Construct truth tables to test each of the following formulas for tautologousness
(validity) on the semantics of [,ukasiewicz, Kleene, and Bochvar, respectively (that
is, do three tables for each formula), and record your results by writing 'tautolo-
gous' or'nontautologous'next to each table:
1. Pv-P
2. P-P
3. (P-q;*(-PvQ)
4. TP*P
5. P-TP
6. (P v -P)v (-TP & -T-P)
Exercise 15.2.2
Use truth tables to test each of the following sequents for validity on each of these
three semantics. (Recall that, according to the definition given in Chapter 3 and
still in effect here, a counterexample is a valuation on which all the premises are
true and the conclusion is untrue.) Record your results by writing'valid' or'inva-
lid' next to each table:
p*e,pFe
1
2
& -Q)
P r- -(Q
3
PFPvQ
The motive for moving to multivalued logics is that certain kinds of propositions
seem to be neither true nor false. This leads some logicians to posit a third value:
neutral, possible, indeterminate-or something of the sort. But once we have
added a third truth value, why not a fourth, a fifth, and so on? One way to stem
this regress is to stipulate that propositions that are neither true nor false do not
have some otber truth value; they have no truth value at all. Such propositions are
said to exhibit truth-value gaps. But how, then, could we calculate the truth value
of a complex formula if, for example, some of its component propositions have
no truth value?
The simplest response would be to declare all such formulas truth-valueless.
The result would be just like Bochvar's three-valued semantics, except that in the
truth tables where Bochvar's logic has an 'I', this semanrics would have a blank.
I\/rrorv Noncrassrcar Loctcs 415
Or one could create truth tables like those for other forms of three-valued logic.
But this approach is hardly novel, differing in no essential respect from three-
valued logics themselves.
A more interesting method-and one that, unlike most multivalued logics,
preserves the validity of all classical inference patterns-is the technique of super-
valuations invented by Bas van Fraassen.6 In propositional logic, a supervalua-
tional semantics assigns to sentence letters the value T, or the value E, or no value
at all. 'We shall call such an assignment a partial valuation. (Note that at one
extreme some partial valuations assign truth values to all the sentence letters of a
formula or set of formulas and, at the other, some assign no truth values at all.)
The truth values of complex formulas, however, are not calculated by truth tables
directly from the truth values of their components. Rather, the calculation takes
into account all of what are called the classical completions of a given partial
valuation.
Let s be any partial valuation. Then a classical completion of s is a classical
ualuation (one that assigns each sentence letter in the relevant formula or set of
formulas one and only one of the values T or F) that fills in all the truth-value gaps
left by s. In other words, a classical completion of : does not change any truth-
value assignment to a sentence letter that has already been made by 5 but merely
supplements assignments made by s, giving each sentence letter of the given for-
mula or set of formulas a truth value. Since each truth-value gap can be filled in by
the assignment either of T or of F, each partial valuation, unless it is classical to
begin with,7 has more than one classical completion. Consider, for example, the
formula '(P v Q) & (R v S)', and let s be the partial valuation of this formula that
assigns the value T to both 'P' and 'R' but no value to 'Q' or 'S'. Then s has with
respect to this formula four classical completions, corresponding to the four ways
of assigning truth value to the sentence letters 'Q' and 'R'. Each classical comple-
tion of s is represented by a horizontal line in the truth table below:
Rい
Rい
Q Q) & V
T
T
T
T
T
T
T
F T
T
T T
F T
T
T
F
T
F
T
F
T
F
F
The columns under 'P' and 'R' list only T's, because these are the values assigned
by the partial valuation J, and they are retained in each classical completion.
To determine the truth value of the compound formula '(P v Q) & (R v S)',
we expand J into a new nonclassical valuation V, called the supervaluation of s.
' See his "singular Terms, Truth Value Gaps, and Free Logic," lournal of Philoso-
phy 63 (1.966):481-95.
7 The definition of a partial valuation given above treats classical valuations as spe-
cial cases of partial valuations. Thus, if; is a partial valuation which is also classi-
cal, the classical completion of s is s itself.
416 CHaprrn 1 5
This is done by calculating the truth value of '(P v Q) & (R v S)' on each of the
classical completions of s, using the valuation rules of classical logic, as in the
table above. If '(P v Q) & (R v S)' is true on each of these classical completions (as
the table shows that it is), then '(P v Q) & (R v S)'is true on the supervaluation
[. If it had been true on none of them, then it would have been false on [,. And if
it had been true on some but not others, then it would have been assigned no truth
value on'[..
A supervaluation, then, is constructed in two stages. First, we define a partial
valuation s, which assigns to each sentence letter the value T, or the value F, or no
value. Next, at the second stage, we construct all the classical completions of s
and use the classical valuation rules to calculate the truth values of complex for-
mulas on each of these classical completions. A formula O (whether atomic or
complex) is then assigned a truth value by the supervaluation I if and only if all
the classical completions of s agree on that truth value; if not, I assigns no value
to O. More formally:
Then both disjuncts 'P' and 'Q' and the disjunction 'P v Q' are truth-valueless on
the corresponding supervaluation "[, because all three formulas are true on some
classical completions of s and false on others. But contrast this valuation with the
partial valuation g of the formula 'P v -P' on which 'P' has no truth value. Here
there are two classical completions of ?-one that makes 'P' true and one that
makes 'P' false. And again both disjuncts are truth-valueless on the corresponding
supervaluation "[ because each disjunct is true on one of the classical completions
of y and false on the other. Yet the disjunction 'P v -P' itself, since it is classically
valid, is true on every classical completion of t. So, although 'P t Q' is truth-
valueless on the supervaluation on which both of its disjuncts are truth-valueless,
'P v -P' is true on the supervaluation on which both of its disjuncts lack truth
value. It follows that disjunction cannot be represented by a truth table in a
supervaluational semantics. The other binary operators are likewise non-truth-
functional.
In a supervaluational semantics we may retain the classical definitions of
such semantic terms as validity and consistencS but we must now understand the
valuations mentioned in their definitions as supervaluations. Thus, for example, a
formula is valid iff it is true on all supervaluations; and a sequenr is valid iff there
is no supervaluation on which its premises are true and its conclusion is untrue.
(And, of course, there are two ways of being untrue: being false and being truth-
valueless.)
SurprisinglS the logic that results from these stipulations is just classical
propositional logic-even though, as we have seen, its semantics differs signifi-
cantly from that of classical logic. Thus, for example, an inference is valid on
supervaluational semantics if and only if it is valid on classical bivalent semantics,
despite the fact that the former, but not the latter, permits truth-value gaps. Let's
prove this:
LetiDr,...,0rY
Then the sequent @r, . . . ,
thatOr,...rO,FYis classi-
valuation on which
Orr. . ., @,areall true
invalid on
there
that @,, . . . , O, are all true
there is a valuation {0r, . . .,
makes@rr...,
but at of whose classical completions
makes by defini-
tion a classical
classical valuation on which
Y is untrue. Thus we have a contradiction.
418 CHnprrn 1 5
iff they are true on all classical completions of s-that is, true no matter what
object in the domain we assign to the denotationless names or what truth value
we assign to truth-valueless sentence letters. Similarly, formulas are false on "[ iff
they are false on all classical completions of s.
This makes most, but not all, atomic formulas containing nondenoting
names truth-valueless. If, for example, a partial valuation assigns to the one-place
predicate 'P' the entire domain as its extension, then for any name ct, even if a is
denotationless, Pcr is true on the corresponding supervaluation, since no matter
which object in the domain we could assign to a, Pa will be true. Similarly, if the
extension of 'P' is { }, then Pa is false, regardless of whether or not cr has denota-
tion. However, if the extension of 'P' contains some but not all of the objects in
the domain, then Pa is truth-valueless when cr lacks denotation, since some assign-
ments of denotation to o would make Po true while others would make it false.
Various generalizations of these results also hold for relational predicates.
All formulas of the form c = o are true on all supervaluations-a pleasing
result. However, where o and B are different names one or both of which are
denotationless, cr = B is truth-valueless, unless the domain contains but one object.
This semantics validates the inferences:
Fa F lxFx
and
VrFr l- Fa
which are invalid on free logics. Indeed, the resulting logic is-despite the truth-
value and denotation gaps-plain old classical predicate logic. The semantics pre-
sented here is, however, only one of several ways to implement supervaluational
predicate logics. Other, generally more complicated, semantic variations yield log-
ics more akin to the free logics of Section 15.1.8
In both propositional and predicate logic, supervaluational semantics allows
us to retain all the inferential structure of classical logic while abandoning the
semantic principle of bivalence. It is thus an interesting compromise between clas-
sical and nonclassical logics.
Exercise 15.3
1. Prove that a formula of propositional logic is classically valid iff it is valid on
supervaluational semantics.
2. Prove that the inferences 'Fa F lrFx' and 'VxFr F Fa' are valid on the superval-
uational semantics for predicate logic sketched in this section.
420
Raorcatv NoncLassrcal Locrcs 421
Early in the sequence of inferences these premises lead to conclusions that are
either wholly true or approximately true. But as they are used to draw conclusion
after conclusion, the conclusions become less and less true so that by the end of
the sequence we arrive at a conclusion that is wholly false.
It is obvious that no particular statement in this sequence is the source of the
error. AII the conditional premises are true) or nearly so. Rather, there is a gradual
progression from truth into error. The reason we cannot locate the error in any
single premise is that the predicate 'is sustainable' is vague. There are paradigm
cases: populations that are clearly sustainable and populations that are clearly
unsustainable. But there are also intermediate cases in which it is "sort of true" but
"sort of false" that the population is sustainable. Our vague notion of sustainabil-
ity defines no sharp boundary at which a population of n is sustainable, but a
population of n + 1, is not. Rather, as the numbers increase it becomes less and less
true that the population is sustainable.
The novelty of this explanation is that it makes truth a matter of degree.
There are, it appears, not just two truth values, but a potential continuum of
values, from wholly false to wholly true. If we take this gradation of truth value
seriously, the result is an infinite valued semantics.
The obvious way to represent these truth values is on a scale from 0 to 1
with 0 being wholly false and 1 being wholly true. Then, in propositional logic,
instead of assigning propositions just the values T or F, a valuation may assign
them any real number from 0 to 1. This, of course, forces us to revise the valuation
rules for the operators. It is by no means obvious how to do this. Several options
are available, as we saw in Section 15.2, when we move from two values to three;
'We
many more present themselves when we contemplate an infinity of values.
shall sample only one infinite-valued semantics-one of the simplest.
Consider, first, the negation operator. If <D is wholly true, then -<D is wholly
untrue and vice versa. Likewise, it seems reasonable to suppose that if <D is three-
quarters true, -(D is only one-quarter true. Thus, as a general rule -<D has all the
truth that (D lacks and vice versa. More formally,
1. y(-o)=1-"r"(o)
This gives negation many of the properties we expect it to have. It follows from
this definition, for example, that the double negation of a formula has the same
truth value as the formula itself. That is,
422 CHAPTER 16
q]r(--Y)
= 1 - y(-Y) By valuation rule 1
= 1-1+Y(Y) Byalgebra
= Y(Y) By algebra
A conjunction would seem to be as true as the least true of its conjuncts. The
truth conditions for conjunctions are best expressed using the notation'min(x, y)'
to indicate the minimum or least of the two values x and y-or the value both 'x'
and'y' express if x = !. Thus, for example, min(0.25, 0.3) = 9.25 and min(1, 1) =
1. The valuation rule for conjunctions, then, is
2. Y@ & Y) = min(fl(o),1/(Y))
Again, we can prove some familiar properties. A conjunction of the form Y & Y,
for example, has the same truth value as Y:
y(Y & Y) = min(Y(Y), .l/(Y)) By valuation rule 2
= Y(Y) By definition of min
Disjunctions are as true as the most true of their disjuncts. This idea may be
expressed by the notation 'max(x, y)', which indicates the maximum or greatest
of the two values x and y, or the value both variables express if x = T
3. Y(O v Y) = max(lr(<D),Y(Y))
'We can now show, for example, that generalizations of De Morgan's laws hold on
this semantics. For example, y(-O & -Y) = Y(-(<D v Y)):
v(-@ & -Y)
: ilixl['-'];,Il;11'- y(y))) BI;:Hmlxl!?
= 1 - max(Y(@),1/(Y)) By algebral
1 - Y(<D v Y) By valuation rule 3
= Y(-(o v Y)) By valuation rule 1.
There are many ways of dealing with conditionals; again we shall choose
one of the simplest. \07e shall assume that a conditional is wholly true if its conse-
quent is at least as true as its antecedent, but that if the consequent is less true than
the antecedent by some amount r, then the conditional is less than wholly true by
that amount. If, for example, y(O) = 0.3 and y(Y) = 0.4, then "f(O .- Y) = 1,
since the degree of truth of the consequent exceeds that of the antecedent. If,
however, the values are reversed, so that y(O) = 0.4 and y(Y) = 0.3, then, since
the antecedent's degree of truth exceeds the consequent's by 0.1, the conditional is
that much less than wholly true; in other words, "[(@ * Y) = 1 - 0.1 = 0.9. These
truth conditions can be expressed by the equation
4. Y@-- Y) = 1 +min(V(@),Y(Y)) - Y(O)
If Y is at least as true as @-that is, Y(<D) < T(Y)-then min(Y(@), y(Y)) - 1/(O)
= 0 and so V(@ Y) = 1. If y((D) > "f(Y), then min(V(o), y(Y)) - y(@) = V(Y)
-
1 The algebraic principle used here is that min(1
- x, 1 - yl = 1 - max(:r, y).
RaorcaLLvNoNclassrcarLocrcs 423
These valuation rules yield familiar systems if we restrict the number of truth
values. If we consider just the values 1 and 0, taking 1 as T and 0 as F, then these
rules define the classical semantics for propositional logic. If we allow a third
valuerVz, and take it to be I, these rules yield Lukasiewicz's three-valued semantics.
In fact Lukasiewicz was the first to state these rules, though he had in mind appli-
cations not to problems of vagueness, but to the problem of future contingents.
Knowledge of these correlations can be used to generate counterexamples to
certain specific theses in the infinite-valued semantics. For example, we may won-
der whether 'P v -P' is valid (true on all valuations), which in the context of fuzzy
logic means having the value 1 on all valuations. We saw, however, in Section 15.2
that this formula is invalid on l-ukasiewicz's semantics, the counterexample being
the case in which 'f('P') = I. Since I translates into the value Yz on the infinite-
valued semantics, the truth-value assignment 1/('P') = % should also produce a
counterexample here. Indeed it does:
It is interesting to note, however, that the truth value of formulas of the form <D v
t/z.These formulas, though not in all instances valid, are
-(D can never drop below
always at least half true!
This does not hold for all classically valid formulas, however. We saw that
the classical tautology '-(P * -P)' is false on Lukasiewicz's three-valued semantics
when P is indeterminate. SimilarlS on the infinite-valued semantics y('-(P * -P)')
= 0 when Y('P') =Y2.
And, as on Lukasiewicz's three-valued semantics, some formulas-those of
the form @ -- @, for example-are valid despite the infinity of values:
In general, formulas which are classically valid may take any truth value on
the infinite-valued semantics, though those of particular forms may be confined to
a particular range of values or even to one value.
I7ith infinitely many truth values available, we also have various choices for
what counts as valid reasoning. We might, for example, define the concept of
validity in either of these ways:
1. A sequent is valid iff there is no valuation on which its premises are
wholly true and its conclusion is not wholly true (i.e., one which the
truth value of the premises is 1 but the truth value of the conclusion is
less than 1).
2. A sequent is valid iff there is no valuation on which the lowest of the
truth values of its premises exceeds the truth value of its conclusion.
Or, generalizing definition 2, we might even make validity itself a matter of degree:
3. A sequent is valid to degree 1 iff it is valid by definition 2; it is valid to
degree 1, - x iff there is no valuation on which the lowest of the truth
values of its premises exceeds the truth value of its conclusion by more
than x, where x > 0.
RaorcarrvNoNclassrcalLocrcs 425
On definition 3, for example, the sequent 'Q F P v -P' is valid to degree /z.For the
highest value'Q'can have is 1, whereas the lowest value'P v -P'can haveis/2.
Thus there is no valuation on which the truth value of the premise exceeds the
truth value of the conclusion by more than /z.Hence this sequent is valid to degree
'L-Vz=1/2.
A sequent is valid to degree 1 according to definition 3 iff it is valid by
definition 2. And any sequent valid by definition 2 is also valid by definition 1,
though the converse does not hold. There are other possible definitions of validity
as well. None is uniquely correct. Rather, on an infinite-valued semantics the con-
cept of validity splinters into an array of concepts.
In particular cases, various concepts of validity may coincide. For example,
the sequent'P & Q F P' is valid on definitions 1 and 2 and hence valid to degree 1
on definition 3. But in other cases, different concepts of validity yield different
results. Consider, for example, modus ponens in the form of the sequent 'P, P * Q
F Q'. While valid on definition 1, this sequent is invalid on definition 2 and valid
only to degree % on definition 3. For on the valuation "[ such that Y('P') = Vz and
V('Q')= ot
: i,:;:"i!'''1")J(Q'D
On Y, therefore, both premises have the value % but the conclusion has the value
0. Thus there is a valuation (namely, Y) on which the lowest of the truth values of
the premises exceeds the truth value of the conclusion, and so the sequent is invalid
by definition 2. This example also shows that the sequent is valid at most to degree
Vz on defrrition 3. To show that it is valid exactly to degree 1/ztakes more work,
but we have already seen what we need to see.
'We
Let's now return to the problem with which we began this section. note
first that conditionals of the form
sequence of such inferences, then, may lead us from near truth to absolute
falsehood.
As this example illustrates, infinite-valued semantics holds considerable
promise for the semantic treatment of at least some kinds of vague inferences. But
it is not wholly satisfactory. Though we might agree, for example, that the truth
value of 'A global population of 6,000,000,001 is sustainable' is slightly less than
the truth value of 'A global population of 6,000,000,000 is sustainable', there is
no fact of the matter as to what these numerical truth values are. Is the truth value
of the first statement, for example,0.3? 0.5? 0.9? And what, exactly, is the numer-
ical difference between the truth values of the two statements?
Part of our inability to answer lies in our ignorance. 'We don't understand
the earth as an ecosystem well enough to know, except within a very broad range,
what populations are sustainable. But part of the problem lies also in the vagueness
of the term 'sustainable'. Even if we knew everything there is to know about the
earth, we would at best be able to offer only artificially precise numbers for truth
values, because the concept of sustainability is essentially vague and qualitative,
lacking any rigorously quantifiable structure.
In the mid-1960s, Lofti Zadeh set out to improve upon infinite-valued se-
mantics by making the truth values themselves imprecise. That is, instead of as-
signing to a statement like 'A global population of 6,000,000,000 is sustainable'
an arbitrarily precise numerical truth value, Zadeh proposed that we assign it an
imprecise range of values. By this he meant not merely an interval of values (say,
the interval between .4 and .5, which itself is a precisely defined entity), but a fuzzy
interval of values. A fuzzy interval is a kind of fuzzy set. And a fuzzy set is a set for
which membership is a matter of degree.
Most concepts,Zadeh argued, define not classical sets, but fuzzy sets. Take
the concept of redness. Some things are wholly and genuinely red. But others are
almost red, somewhat red, only a little bit red, and so on. So, whereas fresh blood
or a red traffic light might be fully a member of the set of all red things, the setting
sun might be, say, halfway a member and a peach only slightly a member. Now in
fuzzy-set theory, membership is assigned strict numerical values from 0 to 1, like
the truth values in infinite-valued semantics. But in defining truth values, Zadeh
compounds the fuzziness. He might, for example, define a truth value AT (almost
true), which rs a fuzzy set of numerical values in which, say, numbers no greater
than 0.5 have membership 0, 0.6 has membership 0.3,0.7 has membership 0.5,
0.9 has membership 0.8, and 0.99 has membership 0.95. Such a fuzzy set of
numerical values is for Zadeh a truth value. A logic whose semantics is based on
such fuzzy truth values is called a luzzylogic.
If infinite-valued semantics presents a bewilderin g array of choices of truth
conditions and semantic concepts) fuzzy logic compounds the complication. Al-
ready in infinite-valued predicate logic, the extensions assigned to predicates must
be fuzzy sets; for if they were classical sets, atomic formulas containing n-place
predicates (n > 0) would always be either true or false. So, for example, the predi-
cate 'is red' has as its extension the ftzzy set of red things described above. Con-
sequently, infinite-valued semantics has the following truth clause for one-place
predicates O and names ct:
Raorcalrv Norcrassrcnr Locrcs 427
Y(@a) = x iff V(cr) is a member of the fuzzy set "f (O) to degree r.
But on Zadeh's fuzzy semantics the extensions of predicates are structures still
more complex than fuzzy sets, structures which, when applied to the extensions of
names, yield fuzzy truth values. The valuation rules for the operators, and the
semantic concepts of validity, consistencn and so on, must all be redefined once
again to accommodate these fuzzy values. In the process these concepts "splinter"
even more wildly than concept of validity does in infinite-valued logic.
Yet despite the complication upon complication, arbitrariness remains. Za-
deh suggests that a fuzzy logic should not employ all possible fuzzy truth values (a
very large set of values indeed!), but only a small finite range of them, and that it
should correlate them with such natural language expressions as 'very true','more
or less true', and so on. But which fuzzy set of numbers corresponds to the English
expression 'more or less true'? And why should we suppose that precisely that set
is what we mean when we say that a particular sentence is more or less true? The
choice of any particular fuzzy value is just as arbitrary as the assignment of a
precise numerical truth value to a vague statement. The arbitrariness does not go
away;it is merely concealed in the complexity.
Arbitrariness notwithstanding, fuzzy logic has found useful application in
artificial intelligence devices. But it has also acquired a certain unwarranted
mystique. Many people are attracted to the idea of a (warm and?) fuzzy logic
because it sounds as if it might offer relief from overtaxing precision. As a result
of this popularity, the term 'fuzzy logic' is often used loosely. In popular science
publications it may mean no more than an infinite-valued logic-or even statistical
or just plain muddle-headed reasoning.
Exercise 16.1
Prove the following metatheorems using infinite-valued semantics:
1. The sequent'P & Q l- P' is valid on definitions 1 and 2 of validity and valid to
degree 1 on definition 3.
2, The sequent'P, P r Q F Q' is valid on definition 1.
3. The formula '-(P & -P)' is not valid (i.e., does not always have the value 1).
16.2:NTUIT:ON:STIC LOGICS
All the logics thus far considered have as their fundamental semantic concept the
notion of truth. Truth is usually understood as a kind of correspondence between
propositions or thoughts and reality. But many philosophers, believing that we
have no access to reality as it is in itself, doubt that we can ever attain truth in this
sense. I may think that the soup is boiling and then go to the stove and see that it
is. In this sense I may confirm my thought that the soup is boiling. But my seeing
or hearing (or even touching) the soup does not, so this line of reasoning goes,
reveal the soup as it is in reality, but only the soup as I see or hear or feel it. I may,
428 CHapren 'l
6
in other words, compare my thought with the soup 4s perceiued by me, but never
with the soup as it is in itself. But if I can never know the relation between my
thought and reality itself, then I can never know truth. And if truth is something
unknowable, then perhaps our semantics should be based on something more
empirical.
One suggestion is to base our semantics on relations not between thought
and reality, but between thought and evidence-relations such as proof, warrant,
or confirmation. My perception of the soup is a form of evidence that proves,
warrants, or confirms my thought or assertion that the soup is boiling. Thus
thought or assertion, which I experience, is compared with evidence, which I also
experience, rather than with reality or the world, which I allegedly never experi-
ence as it is in itself. Another word for such direct experience is'intuition'. Accord-
ingly, the resolve to restrict semantics to entities that can be made evident to direct
experience is called intuitionism.
Intuitionists reject the traditional notion of truth because it posits a world-
as-it-is-in-itself-something we never experience as such-as that to which
thought or assertion corresponds. For the intuitionist this world-in-itself is irrele-
vant to semantics. Having gone this far, one might even be tempted to conclude
that there is no such thing as the world as it is in itself. Some prominent philoso-
phers have drawn this conclusion.
Others have accepted this conclusion only in regard to certain kinds of real-
ity. The originator of intuitionism, L. E. J. Brouwer, was concerned primarily with
mathematics, not with the world in general. He held that mathematical objects
(numbers, functions, sets, etc.) exist only insofar as we construct them or define
the means for their construction. The propositions of mathematics, then, are not
true in the sense of corresponding to some independently existing reality, but
rather merely confirmable, refutable, or neither, by the evidence of our calculations
and proofs.
Lately others-most notably Michael Dummett-have held that even the
semantics of ordinary discourse is best understood as intuitionistic. Dummett's
view is less "subjectivistic" than that of earlier intuitionists. He holds that what
constitutes the meanings of terms must be publicly observable; otherwise, our
ability to learn language would be inexplicable. Thus we learn the meanings of
terms by observing them in use and so associating them with publicly evident
assertibility conditions. Meaning, then, is constituted, not by truth conditions (for
truth, as we know, is not always publicly evident!), but by assertibility conditions.
To know the meaning of a term is to know the publicly evident conditions under
which it is appropriate to assert various sentences containing it.
Regardless of whether intuitionism is confined to a specific domain, or ap-
plied universally, the strategy is always the same: Replace the traditional concept
of truth with some notion of warrant or evidence. This warrant or evidence may
be regarded more or less subjectively, as was the case with the original intuitionists,
or as publicly accessible, as in Dummett.'We shall use the general term confirma-
tion to stand for all of these notions of warrant or evidence. Replacement of truth
by confirmation subtly shifts the meaning of each formula or statement. To assert
a formula O is to say not that it corresponds to reality, but that it is evident,
warranted, or (in our jargon) confirmed. In the domain of mathematics, Brouwer
Raorcarrv Norucrassrcar Loctcs 429
FIGURE 16.1
Possible Semantic Classification of a Statement @ According to
lntuitionism
Confirmation value of (D
@ is confirmed O is unconfirmed
@. But there can be no such table. For suppose the value of <D were "neither," that
is, neither confirmed nor refuted. Then -<D could not have the value "confirmed,"
since that would mean that it is confirmed that (D is refuted and hence that (D itself
is refuted, rather than having the value "neither." But the meaning assigned by the
intuitionist to '-' does not determine which of the other two values ("neither" or
'We
"refuted") -O ought to have. saw in the preceding paragraph that -<D might be
refuted though O is "neither." But it might also be "neither" if O is "neither"l that
is, both the proposition O that Napoleon ate breakfast on Septemb er 9 , 1,807 , and
its negation may be neither confirmed nor refuted. Thus when (D is "neither," -@
may be either refuted or "neither." It follows that intuitionist propositional logic
is non-truth-functional (or, more accurateln non-confirmation-functional). That
is, the semantic values assigned to a complex formula in a given situation cannot
in general be calculated merely from the values of its subformulas in that situation.
Thus no confirmation-functional three-valued logic will suffice.
The most natural formal semantics for intuitionistic logic is in fact biualent
(two-valued). lWe shall think of the values as "confirmed" and "unconfirmed" and
abbreviate them 'C' and 'U', though in the literature they are usually written as'T'
and 'F' or '1' and '0'.
Refutation, on this semantics, is merely a subspecies of nonconfirmation.
More specifically, <D is refuted if and only if -@ is confirmed-in which case (D has
the value U and -@ the value C. @ may, however, have the value U even if -@ also
has the value U. Figure 16.1 illustrates the relations of these concepts.
The particular version of this bivalent semantics which I shall present here
was developed by Saul Kripke.s It has illuminating connections to Kripkean se-
mantics for modal logic-in particular, for the modal system 54. If the material on
Kripkean semantics in Section 12.1 is not fresh in your mind, it might be helpful
to review it before reading on.
A Kripke model V for intuitionistic logic, like Kripkean models for modal
logics, specifies a set,tl/r of "possible worlds" and an accessibility relation -,f on
those worlds. However, the modality represented by the model is not alethic pos-
sibilitn but a specific kind of epistemic possibility-that is, possibility relative to
what has been confirmed. If we think of possible worlds as worlds, then their
possibility consists in the fact that everything that is the case within them is consis-
tent with the available evidence.
But this is not the most illuminating way to think of possible worlds. Indeed,
the central idea of intuitionism is to avoid positing such worlds-in-themselves and
to formulate a semantics only in terms of the available evidence. It is better, then,
to regard these "worlds" as representations of states of evidence-that is, of the
total evidence available to a person or culture at a given time. In each such state,
propositions are categorized not as true or false, but as confirmed or unconfirmed.
Now a proposition which is confirmed in an evidential state Lu must also be con-
firmed in every evidential state that is epistemically possible with respect to w.For
example, if I have confirmed (by calculation, say) that 472 + 389 = 861, then in
every evidential state that is consistent with my current evidence, 472+ 389 = 861.
It is, in other words, epistemically necessary for me that 472 + 389 = 861. Confir-
mation is tbus a kind of epistemic necessity.
Accordingly, the confirmation of a proposition O in a given evidential state
Lu must be represented semantically not merely by the assignment of the value C
to <D at w, but by the assignment of the value C to @ at w and at all "worlds"
(evidential states) accessible from w.
This is a fundamental difference from the way truth is modeled in Kripkean
semantics for classical modal logic. In classical Kripkean models, if we make a
sentence letter O true in a given world of some model, that does not prevent us
from making @ false in any other world of that model. We may distribute the truth
values for a given sentence letter among worlds in any pattern we like, regardless
of how the accessibility relation is structured. (The same principle applies to the
extensions of predicates in modal predicate logic.) Not so on the Kripkean seman-
tics for intuitionistic logic. Because confirmation, unlike truth, is a kind of neces-
sity, not every distribution of confirmation values is admissible. If we assign to (D
the value C at one "world" w of our model, then we must also assign C to (D at
every "world" z such that w=Ru.
Consequently, as we follow accessibility relations out from a given world,
the set of confirmed propositions can only get larger or, at minimum, remain
constant; it never decreases.
Some accounts of intuitionistic logic suggest that -R be understood tempo-
rally so that accessible evidential states are those we might arrive at through time
as we learn more. But this interpretation is misleading, particularly for nonmath-
ematical applications. What we confirm in the future may contradict what we
confirm in the present, if only because (from a classical standpoint) the world has
changed in the meantime. In reality it may also happen that evidence we obtain in
432 C HAPTER 1 6
the future contradicts our present evidence, even if the relevant facts haven't
changed, because our present evidence is mistaken. (I7hat I think is the boiling of
the soup, for example, may turn out on closer inspection to be only the play of
shadow from a nearby fan.) Kripkean semantics, however, never allows anything
confirmed in an evidential state accessible from w to conflict with anything con-
firmed inw.Therefore 4 is best understood as taking us not to possible future
states of evidence, but to states that, consistent with the evidence we have cur-
rently, we might now enloy if we could now have more evidence.5
Because -,? expresses this specific kind of epistemic possibility (namely, con-
sistency with current evidence), -,? must have a specific structure. It must, first of
all, be reflexive-at least if we suppose (as Kripke does) that evidential states are
noncontradictory, that is, self-consistent.
Moreover, -,? must be transitive. That is, if an evidential state ra, is consistent
with (accessible from) another state w1) and a third state w3 is consistent with wr,
then w, is consistent with wr. This follows from the assumption noted above that
every proposition confirmed in a particular evidential state w is also confirmed in
all evidential states accessible from w. For then if wr.Rw2 and w2.Rw3, all the
propositions confirmed in w, are also confirmed in w, and all of those confirmed
in w, are confirmed in rz.. Hence all of those confirmed in w, are confirmed in zr.
But this means that u. is consistent with the evidential state wl so that (since -(
represents consistency with current evidence) wFRw*
Finally, with respect to predicate logic, we must assign a domain to each
evidential state. Intuitively, this domain represents the objects whose existence we
have confirmed. Since it is epistemically possible for there to be objects whose
existence we have not confirmed, we allow that as we move out from a given
evidential state along a trail of accessibility, the domain may grow. But it may not
diminish, for if the existence of an object is confirmed, then it is not epistemically
possible for that object not to exist. That is, if wr.\wr, then the domain of wrmust
include allthe obyects in the domain of w, and maybe some additional objects as
well. \With these principles in mind we make the following definition:
6 It is, of course, possible to provide explicitly for such future (and also past) states
of evidence by adding tense operators to Kripkean intuitionistic Iogic.
RaorcnrryNor.rcrassrcarLocrcs 433
Now, rather than the truth conditions for complex formulas, we give what we
might call their confirmation conditions. Let @ and Y be any formulas and v any
valuation of <D and Y whose accessibility relation is -.c. Then
1. If @ is a one-place predicate and cr is a name, then V((Do, w) = Ciff Y(cr)
eYlQ, w);
Y((Dct, w)=U iff V(a) t.Y(@,tu).
2. ffA is an a-place predicate (n> 1) and cr,, . . . , o, are names, then
V(<Dct,,. . .: o,,, w)=C iff <Y(ct,), . . . , V(cr,)> eY(Q,w);
V((Dct,, . . . 1 ol,tw) =Uiff <V(ct,), . . ., Y(o,)> tIf(Q,w).
3. If ct and B are names, then'[(cr = 9, w) = C iff V(ct) = 7(P);
=U iff V(ct) + Y(p).
"l/(ct = B, w)
Y(VBO, wl =lJ iff for some world z such that w-Ru, and some
potential name o of some object / in L\,,, V1*,1(@"/p, u) + C.
10. "t/(lPO, w) = Ciff for some potential name 61 of some object tin'j:*,
.l/i*,,1((D"/p,
wl=C;
y(lBO, w) =lJ iff for all potential names o of all objects ,!in es*,T 6,,1
(@"1* w) * C.
The clauses for atomic formulas, identity, conjunction, disjunction, and the exis-
tential quantifier are the same as in Kripkean or Leibnizian modal logic, except
that we have replaced 'T' with 'C' and 'F' with 'U'. But the clauses for negation,
the conditional and the biconditional, and the universal quantifier are novel.
The negation operator expresses not merely nonconfirmation, but refutation.
Refutation is a kind of epistemic impossibility. Intuitively, -(D is confirmed (i.e., (D
is refuted) iff current evidence precludes any possibility of the confirmation of <D.
Formally, -(D is confirmed in a given evidential state w iff no state in which <D is
confirmed is compatible with (accessible from) w. Thus negation is an epistemic
impossibility operator; '-' has the same semantics in intuitionistic logic as'[-'-
or, equivalentlS '' O'-in Kripkean modal logic.
To assert @ - Y is to say that any epistemically possible state that confirms
o also confirms Y. Formally, O -t Y is confirmed in an evidential state w itf in
each evidential state compatible with (accessible from) w, either (D is not confirmed
or Y is confirmed. Except for the replacement of truth by confirmation and false-
hood by nonconfirmation, the semantics for the intuitionistic conditional is the
same as that for the classical strict conditional (see Section 12.3).
The biconditional is in effect a conjunction of strict conditionals. To assert
<D * Y is to say that no epistemically possible state differentiates the two.
The universal quantifier 'V' means "for all epistemically possible objects,"
rather than just "for all objects whose existence has been confirmed." It has the
same semantics as 'trV' in Kripkean modal logic. To assert VBO in intuitionistic
logic, then, is not iust to say that @ is confirmed to apply to all objects whose
existence has been confirmed, but to assert that O has been confirmed to apply to
all objects whose existence is compatible with our current evidential state.
The intuitionistic meanings of conjunctions, disjunctions, and existential
statements are, by contrast, more direct analogues of their classical meanings. To
assert O & Y is to assert that both O and Y are confirmed. To assert O v Y is
to assert that at least one of these disluncts is confirmed. And to assert 3B@ is to
claim confirmation of the existence of an obiect to which (D applies.
This new semanrics requires new definitions for the fundamental semantic
concepts. lWe shall call a sequen, intuitionistically valid iff there is no intuitionistic
Kripke model containing some evidential state in which the sequent's premises are
confirmed and its conclusion is not confirmed. Intuitively, this means that any
evidence that confirmed the premises would also confirm the conclusion. Afor-
mula is intuitionistically valid if it is confirmed in all evidential states in all in-
tuitionistic Kripke models. A formula is intuitionistically inconsistent iff there
is no evidential state in any intuitionistic Kripke model in which it is confirmed.
And so on. 'We now demonstrate formally some of the ideas discussed informally
above:
ReorcerryNorlclnssrclrLocrcs 435
・
L〔 E'TIA■ IIE(〕 REM:・ rhe forinula`P∨ ∼P'is nё tlintuitionisticallyvalid.
γ(lPll;11)=U γ(`P',2)=C II
(Clearly -,f is reflexive and transitive and V meets the conditions on
confirmation value assignments imposed by the definition of a valua-
tion.) Since Yl'P',2) = C and 1.R2, therc is some world u (nameln 2)
such that l-Qu andY('P', u) = C. Thus by valuation rule 4, V('-P', 1) *
C. So, 1) ≠ C, by rulc 6ソ (`P∨ ∼P',1)≠ C.Hcnce`P∨ ∼P'
QED
The model here represents an evidential state (world 1) in which 'P' is neither
confirmed nor refuted, but relative to which it is epistemically possible that 'P' be
confirmed (world 2).
This same evidential state provides a counterexample to the inference from
'--P' to 'P'. In this case, because world 2 holds out the possibility that'P'may yet
be confirmed, it is refuted that 'P' is refuted-that is, '--P' is confirmed-even
though 'P' is unconfirmed. Here is the proof:
The inference from'P' to '--P', however, remains valid, as in classical logic. The
proof is as follows:
435 CHaprrn 16
utl C
T('--P', w) * "[('--P', w) *C, there is by
C,Since
rule 4 some world a such that w.Ru andY('-P',ul =C.And
Y('-P', u) = C, again by rule 4,
since such that
u.Rx,lr(P', x) + C. Now since Y is
u.Ru,
C. However, again since Yis an intultlonlstlc
= C, it follows that Y('P', y) =
w"Ry. Hence in particular, since w4u,
Y('P', u1 = g,
Therefore the sequent'P F --P'
Many classically valid formulas remain valid in intuitionistic logic. 'P - P' is an
example:
Thc fOrmula`P一→P'is
reductio that this ls not the
model Y with
sornc Y such that y('P + P' w
Then by sorne u such that
both y('P' u C ul*
.P +
P' intuitionistically
Though intuitionism dispenses with the law of excluded middle, it retains the
classical principle that anything follows from a contradiction:
`P',〃 )≠
`R∼ P卜 Q'
In intuitionistic natural deduction, the rule EFe that allows us to infer anything
from a formula together with its negation replaces the now-invalid classicai nega--
tion elimination rule: from --(D, infer @. The other narural deduction .ules a.e
unchanged.
Also in intuitionism, we lose the interdefinability of most operators. The
operators '-', '&', 'v', and '-' must all be taken as primitive, though O * y is
definable as (<D
- Y) & (Y - (D). This means, for example, that it is not the case
that O * Y is confirmed iff -<D v Y is. For (D
- Y asserts that any evidential state
in which o is confirmed is also one in which y is confirmed (confirmation of @ is
automatically confirmation of Y). But -<D v Y asserts that we have either confir-
mation of -@ or confirmation of Y. rwe may confirm (D y
-tD v Y.
- without confirming
I said earlier in this section that Kripke's semantics for intuitionistic propo-
sitional logic is illuminatingly connected to his semantics for the modal logic sa-
the classical modal logic determined by the stipulation that the accessibility
relation .€ be reflexive and symmetric. The connection can be seen by observing
that in intuitionistic semantics confirmation is a kind of necessity and refutation a
kind of impossibility. That is, to assert that <D is confirmed is ro assert that <D is
confirmed in all accessible worlds (under the presumption that accessibility is re-
flexive and transitive), and to assert that <D is refuted is ro assert that @ is not
confirmed in any accessible world.
Indeed, any formula <D of intuitionistic logic can be translated into a formula
Y of modal logic such that @ is valid in intuitionistic logic iff Y is valid in 54 and
@ follows intuitionistically from a set of premises iff y follows in 54 from the
translations of those premises. Using ':((D) to stanC for the translation of intuition-
istic formula <D into the language of 54, the following rules stipulate recursively
how the translation may be carried our. For any formulas @ andy:
1. If O is atomic, then r(O) = lO
2.:t(-@)=n-z(@)
3. z(@ & Y) = r(o) & r(Y)
4. tlo-vY)-r((D)vr(Y)
5. {@- Y)=r(z(<D)- r(y))
6. z(@* Y) = n(;(o) * r(Y))
7. r(VBo)= rVBr(o)
8. z(lBo)= IFY(1p)
Comparing the valuation rules and the definition of an intuitionistic valua-
tion to the corresponding rules and definitions of Kripkean semanrics for classical
modal logic makes the rationale for these rules clear. The operators '&', 'v', and
'f ' work essentially the same way in intuitionistic logic as in 54, so the translations
438 Crapren 16
This last formula is not valid in 54. We can via such translations reduce any
problem in intuitionistic logic to an equivalent problem in 54.
In summary, the decision to take confirmation rather than truth as the fun-
damental semantic concept yields a logic with some, but not all, of the features of
classical logic.
EЖ ercise 16.2
Prove that each of the f0110wing sequents is valid on iKripkean intuitionistic
sen■ antics:
1.P&Q卜 P
2.RQ卜 P&Q
3.P,P→ Q卜 Q
4。 P卜 P∨ Q
5。 ∼RP∨ Q卜 Q
6。 P→ Q,Q→ R卜 P→ R
PrOve that the follo、 ving are not intuitionistically valid:
7.P→ Q卜 ∼P∨ Q
8.∼ P→ ∼Q卜 Q→ P
PrOvc that the following are intuitionistically inconsistent:
9. P8ζ ∼P
10。 ∼(P―→P)
RaorcaLLv Norucrassrcar Locrcs 439
Relevance logic (also called relevant logic) is a form of logic that does not count
an inference valid unless its premises are relevant to its conclusion. All the logics
we have considered until now validate irrelevant inferences. In particular, the se-
quent 'R -P F Q' is valid in every system we have surveyed. And though most of
the nonclassical logics we have considered lack some of the valid formulas of
classical logic, still the formulas which are valid in those systems validly follow
from any set of premises, whether relevant or not.
In relevance logic, by contrast, an argument is valid only if there is some
relevant connection between its premises and its conclusion. But the nature of this
connection is disputed. Relevantists generally agree, however, on three things:
everyone is guilty of embezzlement, that bologna is blue, and infinitely many other
absurdities. It would be useful, then, to have for the domain of law and for other
domains that may admit contradictions a logic which allows only relevant conclu-
sions to be validly derived from these contradictions. Unlike most other forms of
logic, this new logic would isolate the consequences of contradictions, preventing
them from "infecting" irrelevant areas of knowledge. Logics according to which
contradictions do not imply all propositions are said to be paraconsistent. Rele-
vance logics are paraconsistent logics.
But we need not hold that there actually are true contradictions to have
reservations about such reasoning. Contradictions are frequently encountered in
hypothetical reasoning, and even there it seems odd, if not worse, to reach a
contradiction and then infer something wholly irrelevant.
It seerns similarly perverse to "prove" a valid formula from an irrelevant
premise. Suppose, for example, that someone challenges us to prove that nothing
is both alive and not alive and that we argue as follows:
Abe Lincoln was truthful.
Nothing is both alive and not alive.
(In symbols this might be'Ta l- -lx(Ax & -Ar)'.) Now the premise is true and the
reasoning is classically valid, so the argument is sound. But to respond this way to
a person who was genuinely puzzled about whether something might be both alive
and not alive would be flippant, to say the least. It is perfectly evident that the
premise of this argument does not prove, support, or provide any evidence for its
'We
conclusion. In that sense, the argument is not valid. seem, then, to have an
intuitive, nonclassical notion of validity. Relevantists hope to formalize that
notion.
Finally, relevantists hold that there is a "natural" conditional that is true only
if its antecedent and consequent are relevantly connected. Using '=' to express
this conditional, the proposition 'Snow is white = Rome is in Italy', for example,
is false because its antecedent and consequent, though both true, are irrelevant to
one another. None of the conditionals we have considered so far, not even Lewis
conditionals, require relevance for their truth so that the relevantist's understand-
ing of if is distinctive. In particular, where 'A' and 'B' express unrelated proposi-
tions, relevantists object to inferences such as:
A,BFAI*B
which is valid for the Lewis conditional,
BFA-B
-AFA-B
which are valid for the material and intuitionistic conditionals, and
IBFA-B
-OA l- A- B
which are valid for all the conditionals we have so far studied, except for the Lewis
conditional and the conditional of Bochvar's multivalued logic. For the condition-
als of relevance logic, none of these sequents are valid.
RaorcaLryNor'rclasStCelLocrcs M1
A great variety of semantics have been offered for relevance logics. Some of
these incorporate a non-truth-functional form of relevant conjunction, called fu-
sion, which seems to have no straightforward natural language equivalent. Indeed,
in some of the most interesting forms of relevance logic, fusion is employed in the
metalanguage to define validity.8 Treatment of the semantics of fusion is, however,
beyond the scope of this book.
Here I will present a simple truth-functional relevance semantics for propo-
sitional logic without conditionals. This semantics can be extended to condition-
als, but not very happily.e No one has yet produced a completely satisfying
semantics for relevance logic as a whole.
The essence of this semantics is a radical rejection of bivalence. Classically a
sequent is valid iff it lacks a counterexample. This leads us to accept such sequents
as 'P, -P F Q' and 'P F Q
- Q' as valid. Relevance logicians find this definition too
permissive. There are two ways to tighten it up. The first is to incorporate some
additional criterion of relevance into the definition of validity. The second is to
retain that definition but liberalize our notion of a counterexample to allow novel
counterexamples capable of invalidating sequents like those just mentioned. The
semantics we shall examine here uses the second method.
A counterexample is a valuation which makes the premises, but not the
conclusion, of a sequent true. But what sort of valuation could make the premises
of 'P, -P l- Q' both true or the conclusion of 'P F Q -- Q' untrue?
Actually, we have already seen valuations on which 'Q * Q' is not true. It is
not true, for example, on the three-valued semantics of either Kleene or Bochvar
when 'Q' has the value I. Thus, in general, a three-valued logic can provide coun-
terexamples to sequents whose conclusions are classical tautologies. But a coun-
terexample to 'P, -P l- Q' would have to make both 'P' and '-P' true. What sort of
valuation could do that?
Both premises could, perhaps, be true if 'P' were both true and false. For in
that case '-P'would also, presumably, be both true and false. Hence both premises
would be true, and both would also be false. Semantics which permit the assign-
ment of both values to a proposirion are called dialethic (literally, "two-truth").
To see how such a semantics might have some practical application in the field of
law, consider this instance of 'P, -P F Q':
Corporation X is liable.
Corporation X is not liable.
Bologna is blue.
To this argument we can imagine the following counrerexample. Suppose that the
legislature has enacted contradictory laws which make it both true and false that
Corporation X is liable. Then since it is true that Corporation X is liable, the first
premise is rrue. And since it is false that Corporation X is liable, the second premise
is true. But the conclusion, let us agree, is not true. Of course the premises are both
false as well. But if we define a counterexample as a situation in which the premises
are true and the conclusion is not true, this situation fits that definition.
Originally, however, in Chapter 1, we characterized an informal counterex-
ample as a coherently conceiuable situation in which the premises are true and the
conclusion untrue. It is not so clear that the situation I have described is coherently
conceivable. In one sense it is. Contradictory legislation can be enacted. But in
another sense it isoot coherently conceivable for a proposition to be both true and
false. Dialethicists in effect propose a liberalization of our notion of coherence so
that we can coherently conceive such contradictions in certain situations. Let's
humor them.
We might fear that their proposal would generate counterexamples every-
where, leading us to a wholesale denial of validity. In fact' however, the new
counterexamples it produces (classical counterexamples still stand) apply only to
sequents that we would generally recognize as irrelevant. Dialethicism does not
invalidate, for example, the sequent'P & Q F P' (simplification). Any truth-value
assignment (whether dialethic or not) that makes the premise true must also make
the conclusion true; there is no counterexample. Of course if 'P' is both true and
false and 'Q' is true, then the premise is both true and false and so is the conclusion.
Hence the premise is true and the conclusion false. But this is still not a counter-
example, for we have defined a (formal) counterexample as a valuation on which
the premises are all true and the conclusion is not true. The conclusion of this
inference might be false when the premise is true-and false as well-but in any
case this conclusion cannot fail to be true when the premise is true. Thus simplifi-
cation remains valid on a dialethic semantics.
For the same reason 'P & -P l- P' is valid, though 'P 6a -P F Q' is not.
Contradictions thus have consequences-but only relevant consequences.
Before going on, we ought to be more explicit about the semantics we are
using. Recall that we are considering only the nonconditional fragment of propo-
sitional logic. Since this logic is purely propositional, a valuation can be merely an
assignment of truth values to sentence letters. But since propositions may receive
either value, neither, or both, it is convenient to think of what is assigned to a
sentence letter as a set of truth values. Any of these four sets may be assigned:
More precisely:
nents of a complex formula @ lack truth value but values of the other parts suffice
to determine the value (or values) of @ by classical truth tables, then that truth
value is what is assigned to @. (This is in effect the procedure of Kleene's multival-
ued semantics, but with truth value-gaps instead of a third truth value.) 'where one
or more components of a formula have both values, then the value(s) of that
formula are calculated classically for both.
For example, suppose that V('P') = {T, F} and V('Q') = F. What is V('P &
Q')? Since'Q'is false (and not true), the classical truth table makes'p & e,
false both in the case in which 'P' is true and in the case in which 'P' is false.
Hence, regardless of whether we combine the falsity of 'Q' with the truth of ,p, or
with the falsity of 'P', we ger the same result: 'P & Q' is false. Therefore "f(,p &
Q')= {F}.
If, however "l/('P') : {l F} and "l/('Q') = T then "l/('P & Q') = {T, F}. For in
combining the truth of 'Q'with the truth of 'P'we see that'P & Q' is true, and in
combining the truth of 'Q'with the falsity of 'P'we see that'P & Q' is false.
The following valuation rules express these principles in full generality:
1. TeY(-@)iffFsV(o).
F eY(-<D)iff T eV(<D).
2. T el/(o 6. Y) iff T e"lr(O) and T eY(Y).
F e Y(@ & Y) iff F e T(O) or F e 1/(Y), or both.
3. T eV(Ov Y) iff T eV(O) or TeT(Y), or both.
F eV(@ v Y) iff F e Y((D) and F eV(Y).
These rules may also be represented, though less compactly, as four-valued truth
tables:
司﹁■l
F^
η司■†珊司司司工司■
(T}
{T}
T ︵ や
F ^ F ^ F ^
{T}
{T}
I■IT
{F}
{F}
い^ ﹁
{F}
I司■
{F}
{■ F}
{■ F}
r
{■ F} (continued)
444 CHaprun 16
{■ F} {} {FI
{T} {}
{F} {F}
{■ F} {F}
{} {}
﹁
η
T
η
T
η
T
I司■lη珂■lη町■1
I
η
F
町
司
r
r
■
F
I
l
F
I
I
F
F■ r
■
F
F
r
F
I
T
I
I
F
I
†
Notice that where @ and Y have exactly one truth value each, these are iust the
classical truth tables, and that in other cases these tables retain as much as possible
of the classical valuation rules.
Just as bivalent truth tables can be used to establish the validity or invalidity
of sequents in classical propositional logic, so these tables can be used to demon-
strate the validity or invalidity of sequents in the nonconditional fragment of
'We
propositional relevance logic. simply construct the table according to the val-
uation rules or truth tables for the operators and then scan it, looking for a line on
which the premises are all true (whether or not they are also false) and the conclu-
sion is not true (whether or not it is false). The following table, for example,
demonstrates the validity of one version of De Morgan's laws (in fact, all versions
of De Morgan's laws are valid on this semantics):
P Q P &
p^
Q ∨ ト (∼
∼Q)
﹁﹁T
ηηηηη司■︱
﹁司■1﹁ 司■︱
司 I
T
{F} {F}
F 司 司 司 T ■ l
T
{T}
lη﹁I■︱
{F}
T
F} {F} {■ F}
T
{F} {}
F
I
(T}
F
{F}
I
司
{T} {T}
F^
Fユ
I
F
F} {T} 1■ F}
司
{T} (}
Raorcarrv Norucrassrcar Locrcs 445
司 ﹁ ■ 1司T
町 ■ ■ 司 到 ︱ 司 ︱
F
{■ F} {T} {T} {T}
■ ■ ■
F
F
F 〔Fl {■ F} {■ F}
F ^ ¨
{■ F} {ェ F} {■ F}
F
F
{} {T} {T}
{T} {T} {T}
I■ ︱
{F} {} {}
F
{■ F} {T} {T}
{} {} {}
There is no horizontal line on which the premise is true and the conclusion untrue,
and hence no counterexample, though there are lines on which both the premise
and the conclusion have the value {l F}. In fact, the table shows that 'P v Q' and
'-(-P 6a -Q)' are logically equivalent.
Somewhat surprisingln however, disjunctive syllogism fails. The following
table shows that the sequent'P r Q, -P l- Q' is invalid for dialethic relevance logic:
P Q P ∨ Q, ∼P 卜 Q
T﹁ ﹁ η﹁司■︱﹁■■﹁﹁1﹁ 1
I
司F
﹁ 司 ■ lT
T
I
I 町司﹁ηηη■■■■†︱︱†
﹁
司
やr
T
F}
η
ー
I 司 ■ 4l I 司 ■ ︱ ﹁ 司 ■ 1
司
﹁
F
I
司
F F
F
F^
尉
■
司
l
T
F ^ 口■ F
I
F
司
F
F
﹁
r ﹁
F}
F
r
l
T
I
I
F
F^
F〓
■
︱
On the tenth line of the table, the one on which 'P' has the value {l F} and 'Q' the
value F, we see that the premises are both true (and false) but the conclusion is
untrue. This valuation is a counterexample. (So also is the valuation displayed on
the twelfth line).
Informally, it may help to conceive the counterexample once again in legal
terms. Suppose that 'P' stands for 'Corporation X is liable', which has turned out
to be both true and false in the way described above, and suppose that 'Q' stands
for'The Constitution is invalid'. Then since Corporation X is liable, the disjunc-
tion 'Either Corporation X is liable or the Constitution is invalid' is also true. But
since 'Corporation X is liable' is also false, the second premise, 'Corporation X is
not liable', is true. Still this does not mean that the Constitution is invalid!
The invalidity of disjunctive syllogism (DS) may seem a high price for rele-
vance, but in fact if we allow DS and also allow vl, the sequent 'R -P F Q' becomes
provable at once:
M6 CHnprrn 16
1.P A
2. -P A
3. PvQ 1vI
4.Q 2,3 DS
Since any relevance logic must reject this sequent, relevance logicians have gener-
ally regarded DS as invalid.
Some, however, have rejected vI (or both DS and vI) instead. The rejection
of vI is more in keeping with the criterion of relevance suggested in Section 1.3-
that an inference is relevant only if any idea that occurs in the conclusion also
occurs in at least one of the premises-for vI apparently violates this criterion by
introducing a disjunct that may have nothing to do with any of the previously
asserted premises.
To obtain a dialethic relevance logic that satisfies this criterion, we could
adopt the Bochvar idea that the indeterminacy of the part infects the whole (see
Section 15.2)-using the value { }, of course, instead of the value I. Thus, for
example, while on the dialethic semantics given above a disjunction whose dis-
juncts have the values { } and T is true, on this Bochvar-inspired semantics, such
a disjunction has the value i ). This makes vI invalid and, more generally, guar-
antees the invalidity of any sequent whose conclusion contains an "idea" (sen-
tence letter) not occurring in its premises, thereby satisfying the criterion of
Section 1.3. However, the resulting logic, which lacks both vl and DS, is ex-
tremely weak.
This is the problem with relevance logics generally: Those that satisfy some
of our intuitions about relevance tend to be objectionable in other ways. Thus,
though many relevance logics have been invented, none has emerged as clearly
right or best.
Before leaving the topic of relevance logic, we ought to say a word about
relevant conditionals. Relevance logic was initially developed to formalize the idea
of entailment, or relevant implication. An entailment conditional, like a strict
conditional in classical modal logic, expresses a kind of necessary connection be-
tween antecedent and consequent; but for the entailment conditional that connec-
tion must also be relevant. Thus, whereas a strict conditional is automatically true
whenever its antecedent is impossible or its consequent necessary, an entailment
conditional might be false under these conditions if the relevant connection was
lacking.
Because entailment is a kind of necessary connection, many attempts to for-
mulate a semantics for entailment have employed a modified Kripkean possible
worlds approach. The most prominent of these replace the Kripkean accessibility
relation -c with one or two three-place relations on worlds, which are used to
define the truth conditions for the entailment conditional. There are several ways
of doing this, however, and all have their problems.l0
10 Read, Releuant Logic, and Priest and Sylvan, "Simplified Semantics for Basic Rel.
evant Logics," discuss various forms of Kripkean semantics for relevance logics
with entailment. See notes 8 and 9, respectively.
RaorcalLy Norcrassrcar LoGrcs 447
Exercise 16.3.1
Usc truth tablcs to dctcrinine■ 7hcthcr Or nOt thc fO110wing scqucnts are valid on
the dialethic senlantics fOr rclcvance 10gic givcn abOve:
1.P卜 P
2. P卜 P∨ ∼P
3.P卜 Q∨ ∼ Q
4.P卜 P∨ Q
5。 RQ卜 P
Exercise 16.3.2
Create dialethic truth tables for the operators '-' ,'k' , and 'v' to implement Boch-
var's idea (mentioned above) that the indeterminacy of the parr infects the whole
(i.e., that if a component of a formula has the value { }, so does the entire formula).
Then test the five sequents of Exercise 16.3.1with these new rruth tables.
But suppose assumption 2 was just a presumption which we later find out to be
false. Then we have a new piece of information, which contradicts it:
" Unlike machine languages, such as the language of the ABACUS programs de-
scribed in Sections 10.1-10.3, programming languages like PROLOG do not give
specific directions as to what changes to make in which registers or storage units
Reprcarrv Norucrnssrcnr Locrcs M9
in a particular kind of machine. Because they ignore such specifics, they may be
used on computers of many different designs. However, a programming language
must 6rst be translated into the machine language for the particular kind of ma-
chine on which it is to be used. This is accomplished by a machine-specific rrans-
lating program called a compiler. The compiler in effect tells the computer how to
carry out the details of the tasks the programmer speci6es in the programming
language.
1'z In PROLOG, however, the universal quantifiers are implicit-that is, there are no
symbols for them-and there are no existential quantifiers.
450 (lHAPTER 1 6
offer-is-valid :-
not( offer-is-prohibited ),
not(offer-is-taxed ) .
Conjunctions need no brackets, since all ways of bracketing them are equivalent
so that there is no possibility of ambiguity.
I said that':-' is a generalized conditional operator, because although it func-
tions as a conditional when provided with an antecedent and a consequent, the
antecedent, or the consequent, or both may be omitted. To see what ':-' means in
these cases and to understand how all the meanings are related, it is helpful to
think of a conditional of the form <D i- Y,, . . . , Y, as the classically equivalent
disjunction(Dv -Y, v... v -Y,."The formulai-Yr,. ..,Y* then, is justthis
disjunction without the first disjunct-that is, -Y, v . . . v -Y,. Consequently,
i-Yr,...,Y,assertsthatatleastoneofYr,...rY,isfalse-thatis,thatnotall
ofYr,. ..,Ynaretrue.Similarly,<D:[email protected],withalldisfuncts
but the first missing. It is therefore simply a redundant way of asserting <D. The
clause ':-' is a "disjunction" with no disiuncts; it is called the empty clause. Since
there are no disjuncts, there is no way for this "disjunction" to be true, and so the
empty clause is a contradiction; it invariably receives the value F.
'We
define a literal as an atom or an expression of the form not(@), where (D
is an atom. A formula of PROLOG is known as a clause:
DEFINITION If Y1, . . . ,Y, (n > 0) and @ are literals, the following are
clauses:
i-Yt'"''Y'
@:-Yrr...,Y,
Nothing else counts as a clause.
This definition is PROLOG's sole formation rule. The notation 'Y,, . . . , Y,'
stands for a sequence of literals separated by commas, that is, an a-membered
conjunction-or a single literal in the case n = 1..|n the case z = 0, this definition
makes the symbol ':-' and all formulas of the form (D :- clauses. ':-', as we noted
earlier, is the empty clause. Clauses of the form (D :- may simply be written as <D,
since in that case ':-', as we also noted earlier, is redundant.
!7here O:-Y,, ...,Y,is a clause, @ is called the headandY1,..., Y,the
tail of the clause. Some clauses have no head, some have no tail, some have neither
a head nor a tail, and some have both.
Notice that conditionals may not occur as subformulas of conditionals or
conjunctions and that only atoms may be negated. Even with the language thus
13 Notice that I am using the notation of propositional logic here, not the notation
of PROLOG, though I am omitting unnecessary brackets; PROLOG does not
contain the operators'v' and'-'.
Raotcalrv Norucrassrcal Locrcs 451
'(:―
Ψl,...,Ψ ″ )=Ti“ γ(Ψ l)≠ T Or. .Or γ(Ψ ″
)≠ ■
(:― Ψ l,...,Ψ ″)=F iff γ(Ψ l)=T and ..and γ(Ψ ″)=T
(Φ )=T iff γ(Φ )=T;
:―
'(Φ
:―
)=F iffγ (Φ )≠ T
4. ″
(`:― ')=R
D
'(nOt(Ψ
))=Ti“ γ(Ψ )≠ T;
(not(Ψ ))=F iff γ(Ψ )=T
1a In most actual implementations what we are calling the program is divided into
the program proper and the database. The program proper consists mainly of
conditionals (clauses with both heads and tails) that represent more or less endur-
ing principles, whereas the database consists mainly of atoms (clauses with heads
but no tails) that represent the facts or the state of a computation at the moment.
Usually databases are more fluid, being updated frequently. The distinction be-
tween the program proper and the database is, however, pragmatic, not logical
(from a logical point of view, both consist of program clauses functioning as cate-
gorical assumptions), and we will take no further notice of it here.
452 Cuapre n 'l
6
LOG more expressive power for practical applications, and this accounts for the
partial redundancy.
PROLOG systems incorporate a proof procedure called a resolution algo-
rithm to check arguments for validity. These arguments are all of a particular type.
Their premises are a program, and their conclusion is a query to be tested against
that program. (To keep things simple, we shall consider only atomic queries.) The
resolution algorithm aims to construct a proof of the query from the program. The
strategy is always reductio. Hence, where <D is the query, the proof takes the
program clauses as assumptions and hypothesizes the goal clause :- @, which is in
effect the negation of @ (see valuation rule 2 in the case n = 1). To these clauses,
the algorithm now applies certain inference rules. If the application of these rules
produces the empty clause (a contradiction), the procedure classifies the conclu-
sion as true by reductio. If, however, there comes a point at which all possible
moves with the rules have been made and the empty clause has not been deduced,
then the procedure classifies the conclusion as false. It may also happen that the
rules can be applied infinitely.ls Propositional PROLOG has only three inference
rules. The first of these is just a notational variant of the classical -I. \7e shall call
it reductio ad absurdum:
To how the resolution rule works, we shall consider a small PROLOG pro-
see
gram, which might be a fragment of a tree identification program. This program
consists of four clauses:16
tree-is-a-j ack-pine :-
ri The program consisting of the single clause 'p :- p', for example, fails to halt
when given the query 'p'. PROLOG's resolution algorithm is therefore not a deci-
sion procedure, though certain closely related algorithms are. The availability of
potentially infinite procedures makes PROLOG a more flexible programming
tool than it would have been otherwise.
16 In a real implementation, the first two of these clauses would belong to the pro-
gram proper (see footnote 1 1 ), whereas the second two would probably belong to
the database, representing, for examp[e, our current observations.
RaorcalLv Norcrassrcnr Loctcs 453
tree-has-needles-4-cmlong,
needles-are-divergent
needles-are-divergent :-
needles-spread-away-from-one-another
needles-spread-away-from-one-another :-
tree-has-needles-4-cm-long :-
1. treejs-a-jack-pine :-
tree-has-needle s-4-crn-l ong,
needles-are-divergent A
2. needles-are-divergent :-
needles-spre ad-away-from-one-another A
3. needles-spread-away-frorn-one-another:- A
4. tree-has-needles-4-crn-long:- A
5. I :- tree-is-a-jack-pine H
Now in a PROLOG proof, goal clauses direct strategy. So we look to line 5, the
only goal clause, to find out how to proceed. The resolution rule says that given a
goal clause of the form :- @,Y,, . . . ,Y,and a program clause of the form @ :- @r,
. . . , @*(z ) 0 and m> O),we may infer the goal clause i- Or, . . . , @* Y,, . . . , Y,.
Now the clause at line 5 is of the form :- @,Yr, . . . , Y * where O = 'treejs-a-jaclc
pine' and a = 0. And the clause at line 1 is of the form (D t- @r, . . ., @*where @, =
'tree-has-needles-4-crn-long', @, = 'needles-are-divergent', and m = 2. Thus we
may infer i-Or, ., ., @*,Y,, . . ., Y,, which in this case (since m =2 anda = 0) is
just :- @,,@2-that is,
6. | :- tree-has-needles-4-cm-long,
I needles-are-divergent 1,5 R
1. treejs-a-iack-pine :-
tree-ha s-needles-4-crn-long,
A A A A H
needles-are-divergent
2. needles-are-divergent:-
needles-spre ad-aw ay
-from-one-another
う0
needles-spread-away-from-one-another :-
´” ”
tree-has-needles-4-cm-long :-
う
:- tree-is-a-jack-pine
/0
:- tree-has-needles-4-crn-long,
needles-are-divergent 1,5R
7. :- needles-are-divergent 4,6R
8. :- needles-spread away-frorn-one-another 2,7R
9. 3,8R
10. tree-is-a-jaclcpine :- 5-9 RAA
The empty clause (a contradiction) emerges at line 9, allowing us to reject the
hypothesis at line 10. To grasp the syntactic features of this final step, notice rhar
the goal clause at line 8 is of the form :- O,Y1, . . . ,Y. where O = 'needles-spread-
away-frorn-one-another' and n = 0. Similarly, the program clause at line 3 is of
the formO:-O,, . . .,@-,wherem =0. Thuswemayinfer:-O,,. . ., @-Y,,. ..,
Y,, which in this case (since m = 0 and z = 0) is just ':-'.
So far, all of this is iust meaningless symbol manipulation. But its meaning
willcome alive when we give it a semantics. ActuallS PROLOG proofs may use-
fully be interpreted from either of two distinct semantical perspectives: the logical
semantics defined in part by the valuation rules above, and a procedural semantics,
according to which the clauses in the proof are not declarative statements, but
rather instructions to the computer. 'We shall employ both perspectives, though
our account of the procedural semantics will be very informal.
\We begin with the logical semantics. Here it will be
useful to translate the
proof into the more familiar notation of propositional logic, using the following
interpretation scheme:
J tree-is-a-jack-pine
L tree-has-needles-4-crn-long
D needles-are-divergent
S needles-spread away-frorn-one-another
Rather than treating PROLOG conditionals directly as conditionals, we shall
translate them into equivalent disjunctions (see valuation rule 1). The following
proof in propositional logic corresponds step-for-step ro the PROLOG proof
above:
A A A A H l,
1
J∨ (∼ L∨ ∼D)
2
Dv∼ S
3
S
4
L
5
∼
J br∼ I)
6
∼L∨ ∼D
ξυ
Rnorcarry Norucrassrcar Locrcs 455
6
7. -D
,
2
7
8. -S
,
3
8
9. S&-S &I
9
5
10. and E
一
This, understood classically, is what the PROLOG proof means from a logical
point of view. Lines 6-8 are all obtained by variants of disjunctive syllogism (DS)
that are obviously valid. (In fact, in this format the resolution rule can be seen as a
kind of generalization of DS, though some of its instances are a good bit more
complex than DS itself. R also serves the function of &l at line 9!) Thus from the
logical perspective, this PROLOG proof is just an ordinary derivation in classical
propositional logic, despite the unfamiliarity of the resolution rule.
Of course we can't quite be satisfied with this. The fact that the resolution
rule produces a valid derivation in this case doesn't show that it is always valid.
('Valid' here has its usual classical meaning.) To have confidence in a rule, espe-
cially a rule of this complexity, we need to prove that it is valid. So we shall:
The resolution rule is therefore just as legitimate as modus ponens. Its complexity
makes it a bit of a nuisance for humans to use, but it is an efficient rule for machine
computation.
So much for the logical perspective. Now let's reconsider the same proof (the
PROLOG original, not the translation) procedurally, as the running of a computer
program. From this perspective it takes on a wholly different meaning. A goal
clause, for example, instead of being read as a negative declarative statement, is an
instruction to the computer to pursue the goal of proving the first literal to the
right of the ':-' symbol. (The literals to the right of this first literal, if any, represent
456 CHAPTER 1 6
a succession of such proof goals, to be attempted later if this first goal succeeds.)
That is why goal clauses are called goal clauses.Indeed, on the procedural seman-
tics, we may read the symbol ':-' prefixing a goal clause as the English phrase 'try
to prove'. Thus, for example, the clause on line 5 instructs the computer to try to
prove that the tree is a iack pine. The computer responds by scanning down the
existing proof for clauses whose head is 'tree-is-a-jack-pine'. Line 1 contains
such a clause.
Now a program clause of the form (D :- Y,, . . . , Y ,, such as is found on line
1, is on the procedural interpretation the following conditional instruction: To
prove O, prove Y,, . . . , Y,. Line 1 therefore tells the computer that it could prove
that the tree is a jack pine if it could show that the tree has needles 4 cm long and
that these needles are divergent. Thus proving these two things becomes the com-
puter's intermediate goal at line 6. Once again the computer scans down the exist-
ing proof and determines that at line 4 it is given that the tree has needles 4 cm
long. The first of the two goals has thus been attained. Hence at line 7 the com-
puter deletes this first goal, leaving only the second: to prove that the needles are
divergent. It scans down the list yet again and finds at line 2 the conditional
instruction: To prove that the needles are divergent, prove that they spread away
from one another. So proving that they spread away from one another becomes
the goal at line 8. Again the computer scans down the proof and notes that at line
3 it is given that the needles spread away from one another. Thus the last goal is
attained, and there is no more work to be done. The empty clause at line 9 may be
read as an instruction to stop and output the result that the query has been proved,
that is, that the tree is a jack pine.
In all of this it may be seen that on the procedural interpretation the resolu-
tion rule is not merely an inference rule, but a device for prioritizing goals and
deleting them once they are achieved.
These two semantic viewpoints, the logical and the procedural, provide
equally correct and mutually illuminating understandings of PROLOG proofs.
Neither uniquely represents what is really going on; they are iust two different
ways of understanding what is from a computational point of view a purely syn-
tactic exercise in symbol manipulation. When dealing with PROLOG programs,
it is useful to keep them both in mind.
PROLOG has a third inference rule, the negation as failure rule, which (since
it is the source of PROLOG's nonmonotonicity) is more interesting than the other
two. To see how it works, we shall return to the problem with which we began
this section.'We shall take as our sole premise the clause
offer-is-valid :-
not( offer-is-prohibited ).
'We
Our query is 'offer-is-valid'. are thus seeking a proof of the sequent 'offer-is-
valid :- not(offer-is-prohibited) l- offer-is-valid'. We will examine the reasoning
first from a procedural point of view.
This sequent represents a query, the conclusion 'offer-is-valid', addressed to
a very simple program consisting of one program clause, the premise. The query
sets the goal of trying to prove 'offer-is-valid', that is, of determining whether the
program makes 'offer-is-valid' true. To do this, the computer scans the program
RaorcallvNor.rcmssrcarLocrcs 457
for instructions. In this case the query matches the head of the one and only pro-
gram clause, which gives the instruction: to prove'offer-is-valid'prove'not(offer-
is-prohibited)'. But how can we prove this? PROLOG's method is to set as its next
goal a proof of 'offer-is-prohibited'. \We shall represent the attempt to prove this
as a hypothetical derivation subsidiary to the main proof. The first step is to
hypothesize ':- offer-is-prohibited', which from a logical point of view amounts
to hypothesizing the negation of 'offer-is-prohibited'. As with the main proof
itself, the strategy of this hypothetical derivation, then, is reductio; we aim to
deduce the contradiction ':-'. From a procedural point of view, however, to hy-
pothesize ':- offer-is-prohibited' is simply to establish the subsidiary goal: prove
'offer-is-prohibited'. If the hypothetical derivation yields a contradiction, this
goal succeeds, which means that the prior goal of proving 'not(offer-is-prohib-
ited)' fails. But if all possible moves are exhausted and we still have not attained a
contradiction, then the subsidiary goal of proving 'offer-is-prohibited' fails. It is
at this point that PROLOG makes its great nonclassical leap, for it counts the
failure to prove a clause @ as a proof of not(O). The mechanism by which it does
so is the negation as failure rule. Use of this rule, in other words, enables us
to deduce 'not(offer-is-prohibited)'. But, as we saw earlier, this suffices to estab-
lish 'offer-is-valid' and thus complete the proof. That is the reasoning in a nut-
shell. Now let's consider it more formally. \We begin with a more explicit statement
of the negation as failure rule:
1. offerjs-valid :- not(offer-is-prohibited) A
2. I :- offer-is-valid H
From a logical point of view, ':- offer-is-va1id' is the negation of 'offer-is-valid'.
From a procedural point of view, it is the instruction "prove 'offer-is-valid'."
':- offer-is-valid' is a goal clause (headless clause). As noted above, the
structure of the goal clauses in a proof determines the strategy. A goal clause must
have one of two forms:
:-@rYr,... rYn
:-not(o),Yl,..., Y,
where @ is an atom, Yr, . . . , Y, are literals, and n 2 0. With clauses of the first
form there is no choice but to use the resolution rule; if the rule is not applicable,
the clause cannot be used. But with clauses of the second form, though we can
still use the resolution rule, a second strategy is available: Hypothesize <D and,
458 Cuprrn 16
1. offer-is-valid :- not(offer-is-prohibited) A
2. :- offer-is-valid H
3. :- not( offer-is-prohibited ) 1,2R
Since ':- not(offerjs-prohibited)' is of the form :- not(O),Yl, . . . ,Y* where n =
0, and since 'not(offer-is-prohibited)' does not match the head of any other
clause, the next step, as explained above, is to hypothesize ':- offer-is-prohibited'
for either RAA or NF:
1. offer-is-valid :- not(offer-is-prohibited) A
2. | :- offer-is-valid H
3. I :- not(offer-is-prohibited) 1, 2 R
4. I I :- offer-is-prohibited H
.We
must now see whether this new hypothesis will yield a contradiction. This
hypothesis is of the form :- (D,Y, ; . . . tY,, where @ is 'offer-is-prohibited' andn=
0. Hence the only rule that might be applicable to it is the resolution rule. But there
is no clause of the form offer-is-prohibited t- @,, . . . , @*, which would be needed
to apply the resolution rule. S7e have also exhausted our options with the other
goal clauses at lines 2 and 3. Therefore the attempt to prove 'offer-is-prohibited'
has failed.
Now the negation as failure rule says that if we have hypothesized :- @ and
failed to deduce ':-' after R and NF have been used as many times as possible, we
should end the hypothetical derivation and infer not(O) :-. In this case (D = 'offer-
is-prohibited'. Thus the next step in the proof is this:
1
:- offer-is-valid H
3
:- not offer-is-prohibited )
( 1,2R
4
I :- offer-is-prohibited H
5
1
offer-is-valid :- not(offer-is-prohibited) A
2
3 I :- offer-is-valid H
| '- not(offer-is-prohibited) 1,2R
4
I t :- offer-is-prohibited H
5
3,5R
7
I have stated this algorithm a bit loosely. In actual implementations there are
strict directions for scanning down the proof to locate the first unused formula
of a given type. These details matter to programmers, but we need not bother
with them.
If you were attentive to the earlier exposition, you might have been bothered
by the factthat I did not annotate hypotheses with a rule name. The algorithm
explains why. IThen we introduce a hypothesis-either when we hypothesize ne-
gation of the conclusion or when we follow step 3 of the algorithm-we may not
know which of the two rules RAA or NF will be used to end the hypothetical
460 CHAPTER 1 6
explicit information to the contrary). Thus, for example, the following sequent
is not provable by PROLOG's logic:
offer-is-prohibited :- A
3
:- offer-is-valid H
4
:- offer-is-prohibited H
6
2,5R
7
not(offer-is-valid) :- 3-7 NF
The offer is now presumed not to be valid, since the attempt to prove that it is
valid (lines 3-7) has failed.
PROLOG may at first seem too arcane for practical application; but once
one gets used to how it handles presumptions, it begins to seem natural. Moreover,
if one can manage to formulate a program so that it implies as true just those
propositions in the relevant field of knowledge that are in fact true, then PRO-
LOG's logic gives the same results as classical logic. But even if not, the results
obtained from a PROLOG program are readily intelligible. And PROLOG has
been used in the construction of much useful software-including the ABACUS
program supplied with this book.
Exercise 16.4
Prove the following sequents, using PROLOG's resolution algorithm (premises are
separated with semicolons to avoid confusion with the conjunctive commas):
1. p:-q,t; s:-t; q:-r,s; t:-; r:- l- p
2. p:-q,not(r); q:-; r:-s F p
3. p:-q,not(r); q:-; r:-s; s:- F not(p)
Is there a uniquely true logic? This book's answer, obviousln is no. Classical logic
has dominated the field, particularly in mathematics, and it has the advantage of
simplicity, but it is not the best logic for all applications. The advent of artificial
intelligence systems has spurred the creation of hundreds of alternative logics for
specific applications. Some, like PROLOG, sacrifice strict logical validity to com-
putational efficiency. Others make use of three or more semantic values, because
462 C HAPTER 1 6
for some applications classical two-valued semantics is too restrictive. (We don't
always know which propositions in our database are true or false, so we can't
always make an accurate two-valued assignment-or we may want to grant these
propositions degrees of truth and falsity.) The pressure toward nonclassical con-
cessions is especially acute when we attempt to design intelligent systems that use
something akin to natural language. Here we encounter in bewildering complexity
problems of relevance, nonsense, semantic paradox, reference failure, and so on,
all of which raise challenges to the classical principle of bivalence. And even on
the most fundamental issues, such as whether a semantics should be based on
confirmation or on truth, there are deep disagreements; intuitionists take the for-
mer view, classicists the latter.
Moreover, there are motives for rejecting classicism that we have not even
mentioned in this book. The weird behavior of the most fundamental particles of
matter, for example, has convinced some physicists that the subatomic world is
best characterized by a form of nonclassical logic called quantum logic.'7 And, to
take a very different example, some researchers in cognitive science believe that
thought should be understood as operating not primarily on propositions, but on
image-like structures. The main difference between a proposition and an image is
that a proposition is composed of definite and discrete units (represented by
names, predicates, logical operators, etc.), whereas an image is continuous in the
sense that its parts may shade and blur off into one another. Inference, in this view,
is not the discrete deduction of one proposition of others, but the continuous
transformation of images into new images. Such transformations, some research-
ers now speculate, may form the subject matter of wholly new logics called contin-
uous logics. This work, however, is still in its infancy.
Despite these manifold developments, some logicians still believe that be-
neath all the novelties, complexities, practicalities, and concessions to our various
inabilities lies a philosophically pure logic, a unique language of ideal reason.
Maybe. But if so, we are far from discerning it. On the contrary, the tendency of
philosophical logic in the past century has been almost entirely toward greater
diversity, rather than greater uniry.
As a result, many logicians, your author among them, have grown mightily
skeptical of the idea of a single true or best logic. Each logic, or at least each
semantically interpreted logic, carries with it its own notion of truth. !7hat could
it mean to talk of the truth of a logic itself? Is there some overarching notion of
truth against which we might compare individual logics and find them more or
less true? If so, no one has articulated it. And what could it mean to say that one
logic is best? Best for what?
Logics are tools. And just as there is no single tool that is simply and unqual-
ifiedly best, so too there is no single logic that is simply and unqualifiedly best.
Best rs always relative to an application. For pounding nails, a hammer is best; for
cutting wood, a saw. Likewise with logics.
17 For a good account of quantum logic, see Maria Luisa Dalla Chiara, "Quantum
Logic," in D. Gabbay and F. Guenthn er, eds., Handbook of Pbilosophical Logic,
Volume III: Ahernatiues to Classical Logic (Dordrecht: D. Reidel, 1985),
pp.427-69.
INDEX
453
464 1NDEX
path,66,74 realism,326
closed,58 recursion, 37
finished, 68,217-278 reductio ad absurdum. See also negation,
open,68 introduction rule
permissibility, 3 57 -365 in metatheorn 1.21.-1.22, 127 -128
permissible symbol sequence, 132 in PROLOG,452
possibility. See a/so possible situation; possible reflexive relation, 340
world register,268
alethic, 307-308, 334-335, 335, 340 register machine. See abacus
degrees of, 352-356 relation, 187,189-190
epistemic, 309, 322-323, 43 I -432 identitS 190
logical, 326-327 ,334-33 5, 343-344 relativity theory, 3182
metaphysical, 326-327, 334-335 relevance, l3-16
physical, 334-335, 33 6, 342-344 relevance logic, 14 -1 6, 439 -447
practical, 334-335 repacking, 11 9
relative, 336-344. See a/so accessibility representative individual, 228 -233, 237
possible situation, 6-13, 47-42, 5 l, 7 5 -7 8, representative instance, 228 -233, 23 5 -237
193-794,310. See also possibility; possible resolution algorithm, 452, 459
world resolution rule,452
possible world, 31,0-327. See also possibility; reverse halting program, 289-296
possible situation rigid designator,3l3
Post, Emil,284 rules of inference. See inference rules
postfix position, 181 Russell, Bertrand, 48, L77 -17 9, 292-293,
potential characterization, 3 88 384
potential name. See name, potential
predicate, 1.62, 169, 19 5 54 logic, 344,347
identity, 174-180 relation to intuitionistic logic, 437 -438
one-place,189 54 rule, 328, 341-342, 345
two-place, 167,789 55 logic, 343-347
zero-place, 169,190 satisfiable formula. See consistent formula
predicate variable, 385 scope,31-33
pre6x position, 181 scope ambiguity, 32
premise, 3 second-order logic, 383-393
implicit,262-265 self-contradiction. See inconsistent formula
premise indicator, 17 semantic trees. See trees, semantic
prepositional phrase, 266 semantics, 25,27 , 39. See also model; valuation;
presumption,448 valuation rules
program. See also abacus for Kripkean modal logic, 334-344
PROLOG,451 for Leibnizian modal logic, 310-325,
program clause,451 334-336,343
programming language, 44 8 n logical,454-461
programming n ot ation, 27 2-27 4 Meinongian, 399-406
PROLOG,447-46L for predicate logic, 188-223
proof, 80, 82-83 procedural, 454-461
strategy, 9 I -70'1., 23 8 -23 9 for PROLOG,4 5l-451
proposition,3-4 for propositional logic, 39-78
propositional attitude, 307-308 for second-order logic, 387-393
propositional calculus, 51 sentence letter, 25
propositional lo gic, 29 sequent, 29. See also argument form; valid
sequenti invalid sequent
QED. See quod erat demonstrandum serial relation, 358
quantifier, 1,62, 190. See also existential series,124
quantifier; universal quantifier set,185-188
exchange rules, 238-240 empty, 185
numerical,'1,77 'since', 17-19
objectual, 195 singleton, 1 86
substitutional, 195 'so',17
quantifier scope ambiguiry 158 'some', 15, 164-166. See also existential
quantum logic,462 quantifier
Quine, V. Y. O.,265n,303n sound argument, 12
quod erat demonstrandum, 114 soundness
quotation marks, 20-21. of inference rules for predicate logic,261.
of inference rules for propositionaI logic, 83,
R. See resolution rule 145, 153-1,s7
RAA. See reductio ad absurdum of tree rules for second order logic, 393
range of a function, 1 81 of tree test for predicate logic,245-255
468 lruoex