MANY-VALUED LOGICS

Routledge Companion to the Philosophy of Language, Article 2.6

Nicholas J.J. Smith
24 December 2009

1 Introduction

A many-valued (aka multiple- or multi-valued) semantics, in the strict sense,

is one which employs more than two truth values; in the loose sense it is one
which countenances more than two truth statuses. So if, for example, we say
that there are only two truth values—True and False—but allow that as well
as possessing the value True and possessing the value False, propositions
may also have a third truth status—possessing neither truth value—then
we have a many-valued semantics in the loose but not the strict sense. A
many-valued logic is one which arises from a many-valued semantics and
does not also arise from any two-valued semantics [Malinowski, 1993, 30].
By a ‘logic’ here we mean either a set of tautologies, or a consequence re-
lation. We can best explain these ideas by considering the case of classical
propositional logic. The language contains the usual basic symbols (proposi-
tional constants p, q, r, . . .; connectives ¬, ∧, ∨, →, ↔; and parentheses) and
well-formed formulas are defined in the standard way. With the language
thus specified—as a set of well-formed formulas—its semantics is then given
in three parts. (i) A model of a logical language consists in a free assign-
ment of semantic values to basic items of the non-logical vocabulary. Here
the basic items of the non-logical vocabulary are the propositional constants.
The appropriate kind of semantic value for a proposition is a truth value,
and so a model of the language consists in a free assignment of truth values
to basic propositions. Two truth values are countenanced: 1 (representing
truth) and 0 (representing falsity). (ii) Rules are presented which determine
a truth value for every proposition of the language, given a model. The most
common way of presenting these rules is via truth tables (Figure 1). Another
way of stating such rules—which will be useful below—is first to introduce
?
functions on the truth values themselves: a unary function ¬ and four binary
? ? ? ?
functions ∧, ∨, → and ↔ (Figure 2). Representing the truth value of α (on
a given model) as [α], we then specify the truth values of compound formu-
las as in Figure 3. Once one becomes familiar with the distinction between
connectives and truth functions, it is customary to use the same symbols for
both and to let context disambiguate. As it generally increases readability, I

1
 α β ¬α α∧β α∨β α→β α↔β
1 1 0 1 1 1 1
1 0 0 1 0 0
0 1 1 0 1 1 0
0 0 0 0 1 1

Figure 1 : Classical truth tables

? ? ? ? ?
¬ ∧ ∨ → ↔
1 1 0 1 1 1 1
1 0 0 1 0 0
0 1 1 0 1 1 0
0 0 0 0 1 1

Figure 2 : Classical truth functions

?
[¬α] = ¬ [α]
?
[α ∧ β] = [α] ∧ [β]
?
[α ∨ β] = [α] ∨ [β]
?
[α → β] = [α] → [β]
?
[α ↔ β] = [α] ↔ [β]

Figure 3 : Rules for assigning truth values to compound propositions

2
shall mostly follow this practice below (i.e. omit the ?’s on truth functions).
(iii) Definitions of tautology and logical consequence are introduced. In this
case, a tautology is a proposition which gets the value 1 on every model (e.g.
p ∨ ¬p, p → p), and a proposition α is a logical consequence of the set of
propositions Γ (written Γ |= α) if, on every model on which every proposition
in Γ has the value 1, α has the value 1 (e.g. {p, p → q} |= q, {p} |= p ∨ q).
Classical logic is then the language just introduced together with either the
set of tautologies, or the consequence relation, just defined. The definition
of a logic in terms of a consequence relation is more powerful, in that once
we have the consequence relation, we can reconstruct the set of tautologies
as the set of propositions α such that ∅ |= α. However sometimes we are
interested only in tautologies—hence we allow that a logic may be specified
just by giving a set of tautologies, without a consequence relation.
As we have just seen, a logic can be specified as the one which arises from
a certain semantics. Logics can also be specified in terms of proofs. Proofs
come in many different forms. Consider, for example, axiomatic proof sys-
tems. Finitely many propositions are taken as axioms. One or more rules of
inference are specified: they take one or more propositions as input, and give
a proposition as output. A formal proof (in a given axiomatic system) is then
defined as a finite list of propositions, each of which is either an axiom, or fol-
lows from earlier propositions in the list by a rule of inference. A proposition
α is said to be a theorem of the system if there is a proof whose last line is
α. There are many known axiom systems which have the property that their
theorems are exactly the tautologies of classical logic. Thus, classical logic is
said to be finitely axiomatizable. We can also define consequence relations
using axiomatic proof procedures, by allowing assumptions in proofs. So
axiomatic proof systems give us an alternative way of characterizing logics.
In §2 we look at systems of many-valued semantics and their associated
logics and in §3 we mention some of the uses to which these systems have been
put. Overall space constraints, together with the judgement that readers
coming to the existing literature from a philosophy of language background
will find it harder to gain an overview of the different kinds of many-valued
systems than to find information on applications of one or other of these
systems to particular topics of interest, led to the decision to devote the bulk
of the available space to §2.

3
2 Systems of Many-Valued Logic

2.1 Three Values

Suppose we take the language of classical propositional logic, and give it
a semantics which countenances a third truth value—which we shall write
as ∗—as well as the classical 1 and 0. Taking a (three-valued) model to
be a free assignment of one of these three values to each basic proposition,
the classical truth tables will no longer determine a truth value for every
proposition of the language. For example, if p has the value ∗ on a given
model, then the tables do not specify a value for compound propositions (e.g.
¬p, p → q) which have p as a component. One way to remedy this situation
is to replace the classical tables with three-valued truth tables. There are
many such tables (see Bolc and Borowik [1992, ch.3] for a survey); we begin
with the three that have played the most prominent role in the literature.
Figure 4 shows the Bochvar (aka Kleene weak) tables [Rescher, 1969, 29–30]
[Kleene, 1952, 334]. The idea here is that a compound whose components

α β ¬α α∧β α∨β α→β α↔β

1 1 0 1 1 1 1
1 * * * * *
1 0 0 1 0 0
* 1 * * * * *
* * * * * *
* 0 * * * *
0 1 1 0 1 1 0
0 * * * * *
0 0 0 0 1 1

Figure 4 : Bochvar (aka Kleene weak) tables

all take classical values takes the value that the classical tables assign, while
if any of its components takes the value ∗, the compound takes the value ∗.
Figure 5 shows the Kleene (strong) tables [Kleene, 1952, 334]. As before, a
compound whose components all take classical values takes the value that
the classical tables assign. As for the case where a component has the value
∗, there are two useful ways of thinking about what the tables dictate. First,
we can think of the ∗ as an unknown classical value. So we suppose the *
is a 1 and calculate the value that the classical table would give, and we
suppose the * is a 0 and calculate the value that the classical table would
give: if we get 1 both times, then that is the value in the new table; if we

4
 α β ¬α α∧β α∨β α→β α↔β
1 1 0 1 1 1 1
1 * * 1 * *
1 0 0 1 0 0
* 1 * * 1 1 *
* * * * * *
* 0 0 * * *
0 1 1 0 1 1 0
0 * 0 * 1 *
0 0 0 0 1 1

Figure 5 : Kleene (strong) tables

get 0 both times, then that is the value in the new table; and if we get 1
once and 0 once, then the value in the new table is ∗. (When we are trying
to determine what value should be in the new table where two components
both have the value ∗, we calculate all four possibilities given by replacing
each * with a 1 or a 0.) Second, we can think of ∗ as lying between 1 and
0 on a scale from more true to less true: ∗ is more true than 0 but less
true than 1. Then we can see the conjunction as taking the least true of
the values of its conjuncts, the disjunction as taking the most true of the
values of its conjuncts, and the conditional and biconditional as defined in
the standard classical way from the other connectives: α → β =df ¬α ∨ β
(or ¬(α ∧ ¬β)) and α ↔ β =df (α → β) ∧ (β → α). Figure 6 shows the
Lukasiewicz tables [Lukasiewicz and Tarski, 1930]. These are exactly like the

α β ¬α α∧β α∨β α→β α↔β

1 1 0 1 1 1 1
1 * * 1 * *
1 0 0 1 0 0
* 1 * * 1 1 *
* * * * 1 1
* 0 0 * * *
0 1 1 0 1 1 0
0 * 0 * 1 *
0 0 0 0 1 1

Figure 6 : Lukasiewicz tables

5
Kleene tables except for the values of α → β and α ↔ β when α and β both
have the value ∗: in the Lukasiewicz tables these values are 1, whereas in the
Kleene tables they are ∗.
We come now to the third stage of presenting a semantics: defining no-
tions of tautology and/or consequence. The standard way of defining these
notions is to specify a subset of the truth values as designated. A tautology
is then a proposition which takes a designated value on every model, and
α is a logical consequence of Γ iff, on every model on which every propo-
sition in Γ has a designated value, α has a designated value. (The earlier
definitions in the classical case emerge from this template by setting 1 as
the only designated value amongst the two classical values 1 and 0. Marking
designated values by underlining, we represent this choice as {0, 1}.) In the
three-valued case, the most obvious choice is to set 1 as the only designated
value: {0, ∗, 1}. However it is also not unreasonable to set both 1 and ∗ as
designated values: {0, ∗, 1}. On the former choice, a tautology is a propo-
sition which is always true (and consequence is a matter of preservation of
truth); on the later choice, a tautology is a proposition which is never false
(and consequence is a matter of preservation of non-falsity). By combin-
ing different choices of designated values with different truth tables, we get
different logics:

B3 : {0, ∗, 1} and Bochvar tables

B03 : {0, ∗, 1} and Bochvar tables

K3 : {0, ∗, 1} and Kleene tables

K03 , aka LP [Priest, 2008, 124]: {0, ∗, 1} and Kleene tables

L3 : {0, ∗, 1} and Lukasiewicz tables

L03 : {0, ∗, 1} and Lukasiewicz tables

These logics differ more or less from classical logic. For a start, it is not
hard to see that any semantics in which 0 is not a designated value and whose
tables agree with the classical tables where only 1’s and 0’s are involved—this
includes all the systems just introduced—will be such that all its tautologies
are classical tautologies. Going the other way, all classical tautologies come
out as tautologies of B03 and K03 (given the original stipulation that a many-
valued logic is one which arises from a many-valued semantics and does not
also arise from any two-valued semantics, this means that these logics—
considered as sets of tautologies (the situation is different when we consider
consequence relations)—are not many-valued logics: for while they do arise

6
from many-valued semantics, they also arise from the classical two-valued
semantics); B3 and K3 , however, have no tautologies at all; L3 has some of the
classical tautologies (e.g. p → p) but not all (e.g. p∨¬p has the value ∗ when p
does); L03 has all the tautologies of L3 and some more besides (e.g. p∨¬p) but
it still does not have all the classical tautologies (e.g. ¬(p → ¬p)∨¬(¬p → p)
has the value 0 when p has the value ∗). The story is different again when it
comes to consequence relations; for example, while B03 and K03 have the same
tautologies as classical logic, they do not have classical consequence relations
(e.g. q is a consequence of {p ∧ ¬p} in classical logic, but not in B03 or K03 :
consider a model on which p has the value ∗—hence so does p ∧ ¬p—and q
has the value 0).

2.2 Finitely Many Values

The three-valued tables of the previous section can all be generalized to the
case where we have any finite number n of truth values. It is convenient to
represent these n values as fractions:
0 1 n−2 n−1
, ,..., ,
n−1 n−1 n−1 n−1
So where n = 5, for example, we have the following values: 0, 41 , 12 , 34 , 1. To
generalize the Bochvar tables, we say that [α] is whatever the classical tables
dictate when all α’s components have the values 0 or 1; and otherwise it is
the ‘middle value’. (Where the number n of values is odd, the middle value
is simply 21 , i.e. half of the top value 1; where n is even, the ‘middle’ value is
n−2
taken to be 2(n−1) , i.e. half of the second-top value [Rescher, 1969, 43–4].) To
generalize the Kleene tables, we use the rules in Figure 7. To generalize the

[¬α] = 1 − [α]
[α ∧ β] = min([α], [β])
[α ∨ β] = max([α], [β])
[α → β] = [¬α ∨ β]
[α ↔ β] = [(α → β) ∧ (β → α)]

Figure 7 : Rules for n-valued Kleene systems

Lukasiewicz tables, we use rules exactly like those for the Kleene systems,
except that we replace the rule for the conditional with the following:

[α → β] = min(1, 1 − [α] + [β])

7
(In the rule for the biconditional in the Lukasiewicz systems, the conditional
is then this Lukasiewicz conditional, not the Kleene conditional.) The idea
here is that if the consequent is at least as true as the antecedent, then
the conditional is completely true, while if the antecedent is truer than the
consequent—and the difference between their truth values is k—then the
conditional has the value 1 − k, i.e. it is k less than fully true. It is routine to
verify that if we set n = 3 then the three sets of rules just given specify the
three sets of tables given in the previous section (with ∗ written as 12 ) and if
we set n = 2 then these rules all specify the classical truth tables.
If we now choose a set of designated values, logics emerge. For example,
taking the Lukasiewicz rules together with 1 as the only designated value
yields the n-valued logics Ln . Lindenbaum showed that the set of tautologies
of Lm is a subset of the set of tautologies of Ln (m, n ≥ 2) just in case n − 1
is a divisor of m − 1 [Lukasiewicz and Tarski, 1930, 48]. So, for example,
the tautologies of L9 are a subset of those of L5 (and likewise of L3 ), because
4 (and 2) is a divisor of 8. Note that it follows that no tautologies can
ever be gained as we increase the number of truth values. The logics Ln
are all finitely axiomatizable (see e.g. Malinowski [1993, 39]), but not every
finitely many-valued logic is; for example, Rescher [1969, 157–9] presents a
three-valued logic which is not finitely axiomatizable (given substitution and
modus ponens as rules of inference).
We turn now to a different strategy for generating n-valued systems, due
to Post [1920, 1921]. Note that in the Kleene and Lukasiewicz (but not
Bochvar) systems, the truth value of ¬α is as far below 1 as the truth value
of α is above 0; that is, the distance between 0 and [α] is the same as the
distance between [¬α] and 1. This treatment of negation requires that there
be a meaningful notion of distance between the truth values of the system.
In Post’s systems, by contrast, the truth values are merely ordered : given
any two truth values, we can say which of them is the truer; but we cannot
compare the distances between different pairs of values. We represent the
truth values of the n-valued Post system as follows:

t1 , t2 , . . . , tn

The ordering of the values is this: ti is less true than tj just in case i < j.
Figure 8 shows the truth functions of this system. So the value of ¬α is the
value immediately after the value of α (in the ordering of the truth values
from least true to most true), except in the case where the value of α is
the top value tn , in which case the value of ¬α is the bottom value t1 . The
rule for disjunction is familiar from Kleene and Lukasiewicz: the value of
the disjunction is the truer of the values of the disjuncts. The remaining

8
 (
ti+1 if i 6= n
¬ti =
t1 if i = n
ti ∨ tj = tmax(i,j)
ti ∧ tj = ¬(¬ti ∨ ¬tj )
ti → tj = ¬ti ∨ tj
ti ↔ tj = (ti → tj ) ∧ (tj → ti )

Figure 8 : Truth functions for n-valued Post systems

connectives are defined in the standard ways from ¬ and ∨. Taking Post’s
rule and setting n = 3 yields a three-valued system different from any of the
systems examined in §2.1; setting n = 2 yields classical logic.
It is a familiar fact that in classical logic, some connectives can be de-
fined in terms of others. Similar kinds of results hold in many-valued logics.
For example, Lukasiewicz took ¬ and → as primitive connectives, and de-
fined the others in terms of them (Figure 9). Similarly, Bochvar took ¬ and

α∨β =df (α → β) → β
α∧β =df ¬(¬α ∨ ¬β)
α↔β =df (α → β) ∧ (β → α)

Figure 9 : Defined connectives in Lukasiewicz systems

∧ as primitive, and Post took ¬ and ∨ as primitive. It is also sometimes

possible to define the connectives of one system within another system. For
example, Kleene conditional is definable in terms of Kleene negation and
disjunction (Figure 7); but Kleene negation and disjunction coincide with
Lukasiewicz negation and disjunction; so Kleene conditional is definable in
terms of Lukasiewicz negation and disjunction. In fact all the Kleene connec-
tives are definable in terms of Lukasiewicz connectives, but not vice versa;
and all the Bochvar connectives are definable in terms of Kleene (and hence
Lukasiewicz) connectives, but not vice versa [Bergmann, 2008, 91–2]. A set
of connectives is said to be functionally complete if we can define all pos-
sible connectives from the connectives in that set. In an n-valued system,
an m-place connective has a truth table with nm rows (one for each possible
assignment of the n truth values to the m component propositions). We
specify an m-place connective by putting a truth value in each row of the ta-
m
ble. Thus there are n(n ) possible m-place connectives. It is well known that
in the classical (two-valued) case, all these connectives can be defined using

9
only the five standard connectives introduced at the outset. (Indeed, we do
not need all of them: just ¬ and ∧ will do, or ¬ and →, etc.) This prop-
erty of functional completeness of (some subset of) the set of five standard
connectives carries over to some, but not all, many-valued logics. In every
n-valued Post system, the set containing ¬ and ∨ is functionally complete.
In no n-valued Lukasiewicz system is the set of five standard connectives
functionally complete—although for some (but not all) n this set is precom-
plete, which means that it is not functionally complete, but becomes so with
the addition of any connective which is not already definable in terms of the
standard five [Urquhart, 2001, 266–8].
In all the n-valued systems that we have considered so far, the set of truth
values is linearly ordered—but it need not be so. For example, Belnap [1977]
considers a system with four truth values—T , F , B and N —ordered as in
Figure 10 (with x < y iff one can get from x to y by following arrows). N

> T _@@
~~~ @@
~~~ @@
~~ @
N `@ B
@@ ~~?
@@ ~
@@ ~~
~~
F

Figure 10 : A lattice of four truth values

and B are distinct truth values, with neither greater than the other—so we
cannot take conjunction as min and disjunction as max. However, every pair
of values has a supremum (a least value that is greater than or equal to both
values in the pair) and an infimum (a greatest value that is less than or equal
to both values in the pair). A partially ordered structure with this property
is called a lattice. In any lattice of truth values, the strategy of defining
conjunction and disjunction as min and max generalizes to defining them as
inf and sup. So, for example, we can take N ∧ B = F and T ∧ B = B, and
N ∨ B = T and N ∨ F = N .
We turn now to a new strategy for generating finitely many-valued sys-
tems, due to Jaśkowski [1936]. Suppose we have an m-valued system X and
an n-valued system Y . We can then form a new, (m × n)-valued system
X × Y by taking the product of these two systems. The truth values of the
product system will be elements of the Cartesian product Xv × Yv , where
Xv (Yv ) is the set of values of X (Y ); that is, they will be pairs whose first

10
element is a value of X and whose second element is a value of Y . Truth
functions of the product system are then specified coordinatewise (Figure 11:
superscripts on truth functions indicate to which system they belong). One
X Y
¬hx, yi = h¬ x, ¬ yi
X Y
hx1 , y1 i ∧ hx2 , y2 i = hx1 ∧ x2 , y1 ∧ y2 i
X Y
hx1 , y1 i ∨ hx2 , y2 i = hx1 ∨ x2 , y1 ∨ y2 i
X Y
hx1 , y1 i → hx2 , y2 i = hx1 → x2 , y1 → y2 i
X Y
hx1 , y1 i ↔ hx2 , y2 i = hx1 ↔ x2 , y1 ↔ y2 i

Figure 11 : Truth functions for product systems

system of particular interest is the product of the classical system with itself.
This system has four values:

h0, 0i, h0, 1i, h1, 0i, h1, 1i

Its truth table for negation (for example) is shown in Figure 12. Note that

α ¬α
h0, 0i h1, 1i
h0, 1i h1, 0i
h1, 0i h0, 1i
h1, 1i h0, 0i

Figure 12 : Negation in the self-product of classical logic

its truth tables for conjunction and disjunction could alternatively be arrived
at by setting hx1 , y1 i ≤ hx2 , y2 i iff (x1 ≤ x2 and y1 ≤ y2 ), and then following
the inf/sup strategy of the previous paragraph. When it comes to choosing
designated for a product system, there are two obvious choices: hx, yi is
designated iff x is designated in system X and/or (choose one) y is designated
in system Y . In a self -product system X × X, both choices yield the same
set of tautologies—which is simply the set of tautologies of X [Rescher, 1969,
101].

2.3 Infinitely Many Values

If we wish to countenance more than finitely many truth values, there are
two particularly natural options: we can take as truth values all the rational

11
numbers between 0 and 1 inclusive, or the real interval [0, 1] (containing all
the real numbers between 0 and 1 inclusive). Lukasiewicz considered both
these options. The rules for assigning truth values to compound propositions
stated in §2.2 carry over unchanged. If we take 1 as the only designated
values, the two resulting logics are named Lℵ0 and Lℵ1 respectively. Interest-
ingly, they both have the same set of tautologies [Rescher, 1969, 38–9].
A semantics which takes [0, 1] as its set of truth values (and treats the
connectives truth-functionally) is called a fuzzy semantics; a logic that arises
from a fuzzy semantics is a fuzzy logic. Amongst philosophers, the best-
known rules for assigning truth values to compound propositions in a fuzzy
semantics are the Zadeh rules; the statement of these rules is exactly the same
as that of the Kleene rules in Figure 7. Amongst logicians, the fuzzy systems
?
of most interest are the t-norm fuzzy logics. Where ∧ is a binary function on
[0, 1] which we are going to use to define conjunction, it is natural to want
?
∧ to satisfy the conditions shown in Figure 13. A binary function on [0, 1]
? ?
x∧y = y∧x
? ? ? ?
(x ∧ y) ∧ z = x ∧ (y ∧ z)
? ?
x1 ≤ x2 implies x1 ∧ y ≤ x2 ∧ y
? ?
y1 ≤ y2 implies x ∧ y1 ≤ x ∧ y2
?
1∧x = x
?
0∧x = 0

Figure 13 : Conditions on t-norms

which satisfies these conditions is called a t-norm. If it is also continuous—

in the usual sense—then it is a continuous t-norm. A t-norm fuzzy logic
is determined by a choice of a continuous t-norm as the truth function for
conjunction: truth functions for the other connectives are then defined in
?
terms of this t-norm. The truth function → for conditional is the residuum
of the t-norm: ?
?
x → y = max{z : x ∧ z ≤ y}
?
and the truth function ¬ for negation is the precomplement of this condi-
tional:
? ?
¬x=x→0
Biconditional is then defined in the standard way using conditional and con-
junction. Disjunction will be discussed below. The most important examples

12
 ?
Lukasiewicz: x ∧ y = max(0, x + y − 1)
?
Gödel: x ∧ y = min(x, y)
?
Product: x∧y =x·y

Figure 14 : Important t-norms

of continuous t-norms are shown in Figure 14. These three are fundamental
in the sense that every continuous t-norm is a combination of them [Hájek,
1998, 32]. Note that the Gödel t-norm is the min operation used to define
conjunction in Zadeh/Kleene logic; it is the only idempotent t-norm (i.e. one
?
which satisfies the condition x ∧ x = x) [Klir and Yuan, 1995, 63]. The con-
ditionals and negations derived from these three t-norms (as residuum and
precomplement) are shown in Figure 15. The Lukasiewicz operations are the
(
? 1 if x ≤ y ?
Lukasiewicz: x → y = ¬x=1−x
1 − x + y if x > y
( (
? 1 if x ≤ y ? 1 if x = 0
Gödel: x→y= ¬x=
y if x > y 0 otherwise
( (
? 1 if x ≤ y ? 1 if x = 0
Product: x→y= ¬x=
y/x if x > y 0 otherwise

Figure 15 : Residua and precomplements

familiar ones—which is why the t-norm that gives rise to them is called the
Lukasiewicz t-norm. Note however that this t-norm is not the min opera-
tion used to define conjunction in Lukasiewicz’s many-valued logics (§2.2):
that min operation is the Gödel t-norm, and the conditional and negation to
which it gives rise are not Lukasiewicz’s. The Gödel operations are named
for their discussion in Gödel [1986]. Note that while the Gödel conjunction
is the same as the conjunction in Zadeh/Kleene logic, the conditionals and
negations in these logics are different. The conditional arising from the prod-
uct t-norm was discussed in Goguen [1968–69] and is often referred to as the
Goguen conditional.
It is common to add an additional ‘weak’ conjunction to each of these
fuzzy logics—with the t-norm conjunction then being termed ‘strong’—defined

13
as follows (with subscripts w and s indicating strong and weak):

(α ∧w β) =df α ∧s (α → β)

Whatever continuous t-norm is used for ∧s here (and where → is its residuum),
it turns out that weak conjunction is always the min operation [Hájek, 1998,
36]. Disjunction also comes in strong and weak forms. Weak disjunction is
defined in terms of weak conjunction and conditional:

(α ∨w β) =df ((α → β) → β) ∧w ((β → α) → α)

It always turns out to be the max operation [ibid.]. Strong disjunction is

defined as the dual of strong conjunction:

x ∨s y = 1 − ((1 − x) ∧s (1 − y))

(Where the strong conjunction is a t-norm, this dual operation will always
?
be a t-conorm. A t-conorm is a binary operation ∨ on [0, 1] which satisfies
? ?
conditions exactly like those for a t-norm—i.e. put ∨ for ∧ throughout Figure
?
13—except that we replace the final two conditions with 1 ∨ x = 1 and
?
0 ∨ x = x.) In our three systems, this definition yields the operations shown
in Figure 16. The duality of strong conjunction and disjunction means that

Lukasiewicz: x ∨s y = min(1, x + y)
Gödel: x ∨s y = max(x, y)
Product: x ∨s y = x + y − x · y

Figure 16 : t-conorms/strong disjunctions

where negation is defined as ¬x = 1 − x—for example, in Lukasiewicz t-norm

logic—the de Morgan laws will hold (when stated in terms of negation, strong
conjunction and disjunction).
Hájek [1998] introduces an axiom system for a logic BL (basic logic) which
proves all and only those formulas which come out as tautologies (taking 1
as the only designated value) no matter what continuous t-norm is taken as
the truth function for (strong) conjunction. Each of the three specific t-norm
logics mentioned above is then axiomatized by the addition of an axiom or
axioms—different in each case—to BL.
Thus far we have mentioned only propositional logics. Most many-valued
semantics for propositional logic can be generalized to predicate logic by
following through the following fundamental idea. We identify a subset S

14
of a background set U with its characteristic function fS : a function from
U to the set of truth values. Where x is an object in U and y is a truth
value, fS (x) = y means that x is a member of S to degree y. So the truth
values now function as values or degrees of membership (of objects in sets),
as well as values or degrees of truth (of propositions). An n-place predicate is
assigned an extension as its semantic value: a subset of the set of n-tuples of
members of the domain; that is, a function from this set of n-tuples to the set
of truth values. The truth value of the atomic formula Rab, for example—
comprising a two-place predicate R followed by two names a and b—is then
whatever value the extension of R assigns to the ordered pair comprising the
referent of a followed by the referent of b. In order to give rules for assigning
truth values to universally and existentially quantified formulas, we need to
generalize the conjunction and disjunction operations (respectively) so that
they assign values to sets—not just pairs—of values. For details, see Smith
[2008, §1.2, §2.2].

2.4 Non-Truth-Functional Systems

All the systems we have looked at so far are truth-functional: the truth value
of a compound proposition depends only on the truth values of its compo-
nents; that is, the semantics of the connectives are given by associating them
with truth functions. One notable method for assigning multiple values to
propositions in a non-truth-functional way is that of probability logic, where
values—probabilities—are assigned in accordance with the probability cal-
culus. The value of p ∨ q, for example, does not then depend only on the
values of p and q: it also depends on the content of p and q, and in particular
on whether these propositions are mutually exclusive. Interestingly, the set
of tautologies of probability logic (with 1 as the only designated value) is
the same as the set of tautologies of classical logic [Rescher, 1969, 186–7].
Another notable method is that of supervaluations [van Fraassen, 1966]. We
begin with a three-valued model (i.e. a mapping from basic propositions to
the values 0, 1 and ∗). Say that a classical model M2 extends a three-valued
model M3 iff M2 is exactly like M3 except that where M3 assigns ∗ to a
basic proposition, M2 assigns 1 or 0. The rule for assigning truth values to
compound propositions is then this: [α] = 1 (0) on M3 iff [α] = 1 (0) on
every classical model which extends M3 . So, for example, where p and q
both have the value ∗, p ∧ q will also have the value ∗; yet even though ¬p
will have the same value as q, p ∧ ¬p will have the value 0. So the system
is not truth-functional. A variant of supervaluationism is subvaluationism,
which differs by having ‘some’ in place of ‘every’ in the rule for assigning
truth values to compound propositions: [α] = 1 (0) on M3 iff [α] = 1 (0)

15
on some classical model which extends M3 . If we take 1 as the only desig-
nated value, then the (single-conclusion) consequence relation arising from
the supervaluationist semantics is identical to the classical consequence rela-
tion (see e.g. Smith [2008, 82]), while that arising from the subvaluationist
semantics is not (e.g. adjunction fails: α, β 2 α ∧ β). However, if we move
to a multiple-conclusion consequence relation, then the duality between su-
pervaluationism and subvaluationism is restored: neither yields the classical
consequence relation [Hyde, 1997]. Another variant is the degree-theoretic
form of supervaluationism, which assigns values in the set [0, 1]. We start
with a fuzzy model (i.e. a mapping from basic propositions to [0, 1]), intro-
duce a measure on the set of classical extensions of this fuzzy model, and
then say that the degree of truth of a compound sentence on the fuzzy model
is equal to the measure of the set of classical extensions on which it is true.
For further details see Smith [2008, §2.4.1].

2.5 Many-Valued Systems in the Loose Sense

Some of the motivations for many-valued systems to be discussed in §3 lead
most naturally not to the idea of additional truth values, alongside the clas-
sical 0 and 1, but to the idea of propositions possessing neither (or both) of
the classical values. This idea leads to what we termed at the outset many-
valued systems in the loose sense. Instead of adding truth values, we stick
with the classical set {0, 1}, but we alter the definition of a model. Instead of
consisting in a function which assigns a truth value to each basic proposition,
it now consists in a relation between the set of basic propositions and the set
of truth values. Where a proposition is related to just one truth value we
have the analogues of classical truth and falsity; where a proposition is re-
lated to neither truth value we have a truth value gap; where a proposition is
related to both truth values we have a truth value glut. Allowing gaps or gluts
but not both yields a three-status system; allowing both yields a four-status
system. The question of how to assign truth values to compound propo-
sitions when some of their components have gaps or gluts or both may be
addressed in ways analogous to any of those already considered in relation to
three- or four-valued systems: no essentially new ideas are involved—simply
the translation of existing ideas into the new setting (e.g. we might employ
truth tables—with empty and/or multiply-filled cells in place of third and/or
fourth values—or we might employ a version of supervaluationism; for fur-
ther details see e.g. Smith [2008, §2.3, §2.4]). Some writers (e.g. Kripke [1975,
700, n.18]) are adamant that there is a deep conceptual difference between
positing a third truth value and allowing truth value gaps, but in practice
it is often more straightforward to formulate systems with additional values

16
and assignment functions rather than two values and assignment relations,
and so even when the guiding idea is many-valued in the loose sense, the
implementation is often many-valued in the strong sense (e.g. Blamey [1986],
Langholm [1988]).

3 Uses of Many-Valued Logics

One of the earliest motivations for many-valued logics—going back to Aris-

totle [Lukasiewicz, 1930, 63–4]—was the consideration of modal propositions
(e.g. ‘It is possible that p’) and propositions concerning contingent future
events (e.g. ‘I shall be in Warsaw at noon tomorrow’). In relation to the
former, Lukasiewicz [1930, 51] saw that the semantics of the operator ‘It is
possible that’ cannot plausibly be given by any of the four possible one-place
two-valued truth functions; in relation to the latter, he argued that assigning
‘I shall be in Warsaw at noon tomorrow’ either 1 or 0 as truth value now
conflicts with the fact that it is possible, but not necessary, that I shall be in
Warsaw at noon tomorrow, and he writes that his “three-valued system of
propositional logic owes its origin to this line of thought” [53]. This was also
an important motivation for Peirce and MacColl (see Rescher [1969, 4–5] for
references).
The consideration of probabilistic propositions, and in some cases propo-
sitions arising from quantum mechanics, was a motivation for MacColl, Za-
wirski, Reichenbach and Carnap (see Rescher [1969, 14–5, 184–8, 210–1] and
Malinowski [1993, 66] for references).
Consideration of the paradoxes—set-theoretic (e.g. Russell’s) and/or se-
mantic (e.g. the Liar, where it seems impossible to assign either truth value
1 or 0 to ‘This sentence is false’)—was a motivation for Bochvar, Moh Shaw-
Kwei and others (see Rescher [1969, 13, 29, 207] for references).
Kleene was motivated by a consideration of partial recursive predicates.
Think of such a predicate as coming with an algorithm. Given some objects as
input, the algorithm terminates with the answer ‘Yes’: the predicate is true of
these objects. Given some (other) objects as input, the algorithm terminates
with the answer ‘No’: the predicate is false of these objects. But for some
(other) objects, the algorithm does not terminate with either answer: the
predicate is undefined for these objects. Hence Kleene’s three truth values:
t (true), f (false) and u (undefined). A related application of many-valued
logics is in the analysis of category mistakes arising from predicates which are
truly applicable to some objects, falsely applicable to others, and inapplicable
to yet other objects [Rescher, 1969, 28].
Kleene [1952, 335] also considers a different interpretation of his three
values: “t, f, u must be susceptible of another meaning besides (i) ‘true’,

17
‘false’, ‘undefined’, namely (ii) ‘true’, ‘false’, ‘unknown (or value immaterial)’.
Here ‘unknown’ is a category into which we can regard any proposition as
falling, whose value we either do not know or choose for the moment to
disregard; and it does not then exclude the other two possibilities ‘true’ and
‘false’.” This kind of epistemic application of many-valued logics is also
found, for example, in Belnap [1977].
Many-valued logics have important applications in proof theory. For ex-
ample, they can be used to show that certain axioms in a given axiomati-
zation (e.g. of classical propositional logic) are independent of others. This
application originates with Bernays; for illustrative examples see Malinowski
[1993, 105–6] and Mendelson [1997, 43–5]. They can also be used for showing
the consistency of a set of axioms (i.e. that not everything can be deduced
from it). This application originates with Post [Rescher, 1969, 106].
Like classical logic, many-valued logics have applications in circuit/switch-
ing theory; for discussion and references see Malinowski [1993, 109–10] and
Hähnle [2001, §8]. They have also been applied to scheduling problems [Ack-
ermann, 1967, 78]. Many-valued logics have been employed in automated
theorem-proving; for a discussion of the many further applications of many-
valued theorem provers see Hähnle [1993, ch.7].
Many-valued logics have been applied to the analysis of presupposition.
‘Bill knows that Ben is late’ presupposes that Ben is late; ‘The king of France
is bald’ presupposes that there is a king of France; and so on. The many-
valued approach to presupposition holds that if one of a sentence’s presuppo-
sitions is false, then the sentence is neither true nor false. This is known as
a semantic approach to presupposition, because it holds that presupposition
failure affects the truth value of a sentence. For further details see Gamut
[1991, 178–90, 212–4].
A closely related use of many-valued logics is in the treatment of singular
terms which lack referents (e.g. names such as ‘Pegasus’ and definite descrip-
tions such as ‘the king of France’). On this approach, atomic statements
involving such terms are neither true nor false. This idea, which originated
with Strawson [1950], was the motivation for the development of supervalu-
ationism by van Fraassen [1966].
The final use of many-valued logics which we shall mention here is in the
treatment of vagueness, where the basic thought is that statements about
borderline cases of vague predicates (e.g. ‘Bob is tall’, where Bob is neither
clearly tall nor clearly not tall) are neither true nor false. Supervaluationist
and fuzzy approaches in particular have played an important role in the
vagueness literature (see Smith [2008] for further discussion and references).

18
Related Topics

2.8 Intuitionism; 4.3 Presupposition and Implicature; 4.5 Compositionality;

4.9 Reference, Truth and Fiction; 4.11 Vagueness; 4.14 Empty Names.

References

Robert Ackermann. An Introduction to Many-Valued Logics. Routledge &

Kegan Paul, London, 1967.
Nuel D. Belnap. A useful four-valued logic. In J. Michael Dunn and George
Epstein, editors, Modern Uses of Multiple-Valued Logic, pages 8–37. D.
Reidel, Dordrecht, 1977.
Merrie Bergmann. An Introduction to Many-Valued and Fuzzy Logic: Se-
mantics, Algebras, and Derivation Systems. Cambridge University Press,
Cambridge, 2008.
Stephen Blamey. Partial logic. In D. Gabbay and F. Guenthner, editors,
Handbook of Philosophical Logic, volume III, pages 1–70. D. Reidel, Dor-
drecht, 1986.
Leonard Bolc and Piotr Borowik. Many-Valued Logics, volume 1. Springer-
Verlag, Berlin, 1992.
D.M. Gabbay and F. Guenthner, editors. Handbook of Philosophical Logic,
volume 2. Kluwer, Dordrecht, 2nd edition, 2001.
L.T.F. Gamut. Logic, Language, and Meaning, volume I. University of
Chicago Press, Chicago, 1991.
Kurt Gödel. On the intuitionistic propositional calculus. In Solomon Fefer-
man et al., editor, Kurt Gödel: Collected Works, volume I, pages 222–5.
Oxford University Press, New York, 1986. Originally published in 1932.
Joseph Goguen. The logic of inexact concepts. Synthese, 19:325–73, 1968–69.
Reiner Hähnle. Automated Deduction in Multiple-valued Logics. Clarendon
Press, Oxford, 1993.
Reiner Hähnle. Advanced many-valued logics, 2001. In Gabbay and Guen-
thner [2001], pages 297–395.
Petr Hájek. Metamathematics of Fuzzy Logic. Kluwer Academic Publishers,
Dordrecht, 1998.
Dominic Hyde. From heaps and gaps to heaps of gluts. Mind, 108:641–60,
1997.
Stanislaw Jaśkowski. Investigations into the system of intuitionist logic, 1936.
In McCall [1967], pages 259–63.
Stephen Cole Kleene. Introduction to Metamathematics. D. Van Nostrand,
Princeton, 1952.

19
George J. Klir and Bo Yuan. Fuzzy Sets and Fuzzy Logic: Theory and Ap-
plications. Prentice Hall, Upper Saddle River, NJ, 1995.
Saul Kripke. Outline of a theory of truth. Journal of Philosophy, 72:690–716,
1975.
Tore Langholm. Partiality, Truth and Persistence. CSLI, Stanford, 1988.
Jan Lukasiewicz. Philosophical remarks on many-valued systems of proposi-
tional logic, 1930. In McCall [1967], pages 40–65.
Jan Lukasiewicz and Alfred Tarski. Investigations into the sentential calculus,
1930. In Tarski [1956], pages 38–59.
Grzegorz Malinowski. Many-Valued Logics. Clarendon Press, Oxford, 1993.
Storrs McCall, editor. Polish Logic: 1920–1939. Clarendon Press, Oxford,
1967.
Elliott Mendelson. Introduction to Mathematical Logic. Chapman & Hall,
London, fourth edition, 1997.
E.L. Post. Introduction to a general theory of elementary propositions. Bul-
letin of the American Mathematical Society, 26:437, 1920.
E.L. Post. Introduction to a general theory of elementary propositions. Amer-
ican Journal of Mathematics, 43:163–85, 1921.
Graham Priest. An Introduction to Non-Classical Logic: From Ifs to Is.
Cambridge University Press, Cambridge, second edition, 2008.
Nicholas Rescher. Many-valued Logic. McGraw-Hill, New York, 1969.
Nicholas J.J. Smith. Vagueness and Degrees of Truth. Oxford University
Press, 2008.
P.F. Strawson. On referring. Mind, 59:320–44, 1950.
Alfred Tarski. Logic, Semantics, Metamathematics: Papers from 1923 to
1938. Clarendon Press, Oxford, 1956. Translated by J.H. Woodger.
Alasdair Urquhart. Basic many-valued logic, 2001. In Gabbay and Guenthner
[2001], pages 249–95.
Bas C. van Fraassen. Singular terms, truth-value gaps and free logic. Journal
of Philosophy, 63:481–95, 1966.

Further Reading

For readers whose only background in logic is a typical introductory course or

book: Bergmann [2008] provides a thorough introduction to Bochvar, Kleene
and Lukasiewicz three-valued logics, and fuzzy logics.
For readers with greater background in logic: both Malinowski [1993], and
Urquhart [2001] plus Hähnle [2001], provide compact, wide-ranging surveys of
the field, including important details about consequence relations, functional
completeness, proof theory, decidability and complexity, and other topics of

20
central importance in logic; for t-norm fuzzy logics see Hájek [1998].
For readers wishing to explore applications of many-valued logics with
particular relevance to philosophy of language: Smith [2008] gives an ex-
tended argument for the conclusion that the correct account of vagueness
must involve a many-valued semantics.

Biographical Note

Nicholas Smith is Senior Lecturer in Philosophy at the University of Sydney.

He did his undergraduate work at Sydney and his graduate work at Princeton
(PhD 2001). His book Vagueness and Degrees of Truth was published by
Oxford University Press in November 2008.

Word Count

6,055

21