Modus Tollens Probabilized

CARL G. WAGNER
University of Tennessee, U. S. A.

Abstract

We establish a probabilized version of modus tollens, deriving from
p(E|H) = a and p(Ē) = b the best possible bounds on p(H̄). In particular,
we show that p(H̄) → 1 as a, b → 1, and also as a, b → 0.

Keywords

conditional probability, modus ponens, modus tollens, probability logic

1 Introduction

Given an exact or approximate specification of the probabilities of propositions
A1, . . . , An, what are the narrowest bounds that can be placed on the probability of
some other proposition B? This problem was first investigated by George Boole
[B54], whose work was revived by Theodore Hailperin [H65, H84, H96] in a
sequence of publications that offered a fully rigorous, comprehensive treatment
of the problem from the standpoint of linear programming. We follow Hailperin
in calling this problem Boole’s General Probability Problem.

While the propositions in question may be arbitrary, special interest naturally
attaches to cases in which the inference from the conjunction of the “premisses”
Ai to the “conclusion” B is truth-functionally valid, symbolized in what follows by
A1, . . . , An � B. For solutions to Boole’s General Probability Problem in such cases
may be expected to reveal the degree of confidence in the premisses Ai sufficient
to ensure some desired degree of confidence in B, as measured by their subjective
probabilities. Elementary examples include the valid argument

A1, . . . , An � A1 ∨···∨ An,

which has the probabilistic counterpart

p(Ai) = ai ⇒ max{ai} 6 p(A1 ∨···∨ An) 6 min{1, a1 +···+ an}, (1)

and the valid argument

A1, . . . , An � A1 ∧···∧ An,

1



2 Proceedings in Informatics

which has the probabilistic counterpart

p(Ai) = ai ⇒ a1 +···+ an −(n − 1) 6 p(A1 ∧···∧ An) 6 min{ai}. (2)

The right-hand inequality in (1) and the left-hand inequality in (2) are, respec-
tively, the familiar inequalities of Boole and Bonferroni. That the bounds in (1)
and (2) are the best possible was apparently first shown by Maurice Fréchet [F35].

A number of valid arguments of propositional logic involve the truth-function-
al connective −→, variously termed the conditional, the indicative conditional,
and material implication, the truth table of A → B being identical with that of
Ā ∨ B. In the next two sections we consider probabilistic counterparts of modus
ponens (E → H, E � H) and modus tollens (H → E, Ē � H̄). In section 2, a premiss
such as E → H is simply replaced by its probability p(E → H). In section 3, it is
replaced by the arguably more salient conditional probability p(H|E). Our results
on the conditional probability counterpart of modus tollens are new, and exhibit a
feature that is unusual in these sorts of problems, namely, that a high probability
of H̄ is entailed not only when p(E|H) and p(Ē) are high, but also when they are
low.

2 Probabilities of Conditionals

The valid argument E → H, E � H, known as modus ponens or the rule of detach-
ment, has the probabilistic counterpart

p(E → H) = a and p(E) = b ⇒ a + b − 1 6 p(H) 6 a, (3)

for all a, b such that a + b > 1, the bounds in (3) being the best possible. These
results follow immediately from the observations that

p(E → H) = p(Ē ∨ H) 6 p(Ē) + p(H)

and

p(H) = p(E ∧ H) + p(Ē ∧ H) 6 p(E ∧ H) + p(Ē) = P(E → H),

along with easily constructed examples of probabilities p satisfying the antecedent
conditions of (3) and, respectively, (i) p(Ē ∧ H) = 0 and (ii) p(Ē ∧ H̄) = 0. Hail-
perin [H84] has also given an interesting proof using linear programming. Note
that the condition a + b > 1 is necessary and sufficient to ensure that the set of
probabilities satisfying the antecedent conditions of (3) is nonempty.

It follows from (3) that

p(E → H) > 1 − ε and p(E) > 1 − ε ⇒ p(H) > 1 − 2ε (4)

for all ε 6 1/2, as noted by Patrick Suppes [S66].



Carl G. Wagner: Modus Tollens Probabilized 3

From the fact that p(H → E) = p(Ē → H̄) it follows trivially from the above
that the valid argument H → E, Ē � H̄, known as modus tollens, has the proba-
bilistic counterpart

p(H → E) = a and p(Ē) = b ⇒ a + b − 1 6 p(H̄) 6 a, (5)

for all a, b such that a + b > 1. In particular,

p(H → E) > 1 − ε and p(Ē) > 1 − ε ⇒ p(H̄) > 1 − 2ε (6)

for all ε 6 1/2. Understandably, not much is usually made of (5) and (6), since
they are trivial variants of (3) and (4). We include them here only for purposes of
comparison with results in the next section.

3 Conditional Probabilities

3.1 Adam’s Thesis

Truthful speakers do not assert propositions that they do not take to be highly
probable. As David Lewis [L76] puts it, “assertibility goes by subjective prob-
ability.” But Ernest Adams [A65, A75] has pointed out that conditionals seem
to be an exception to this rule, with the assertibility of E → H going not by the
probability p(E → H) of the conditional but, rather, by the conditional probability
p(H|E).

Adam’s evidence consists of a number of cogent examples that accord with
his thesis. Lewis [L76] offers an additional argument: When p(E → H) is high
because p(Ē) is high and, moreover, p(H̄|E) is high, then it is pointless, and in
fact misleading, to assert E → H, rather than Ē. Taking the product p(Ē)p(H̄|E)
as a simple measure of the lessening of assertibility of E → H, we then arrive
at p(E → H) − p(Ē)p(H̄|E), easily shown to be equal to p(H|E), as a suitable
measure of the assertibility of E → H.

Frank Jackson [J79] offers a further argument in support of Adam’s thesis:
Suppose that we have assessed the prior probability p(E → H). Further inquiry
convinces us of the truth of E, and in such a way that the revision of p to the
posterior q(·) = p(·|E) is judged to be reasonable. Then

q(E → H) = p(E → H|E) = p(H|E) and q(E) = 1

and so q(H) > p(H|E) by (3). A large value of p(E → H) is of no utility in such
dynamic cases of modus ponens. What matters is the value of P(H|E).

Even apart from the above discussion, it is clear that conditional probabilities
are salient in a way that probabilities of conditionals are not. For the probability
that H is true, given that E is true, is conceptually simpler and easier to assess
than the probability that E is false or that E and H are both true. Moreover, when



4 Proceedings in Informatics

the probabilities in question are observed or estimated relative frequencies, prob-
abilities of conditionals are of virtually no interest.

For the above reasons it is of interest to investigate probabilistic versions of
valid arguments such as modus ponens and modus tollens in which conditional
probabilities play a central role, a subject to which we turn in what follows.

3.2 Modus Ponens for Conditional Probabilities

Replacing p(E → H) by p(H|E) yields the following variant of (3): If 0 6 a 6 1
and 0 < b 6 1 , then

p(H|E) = a and p(E) = b ⇒ ab 6 p(H) 6 ab + 1 − b, (7)

the bounds in (7) being the best possible. This result follows immediately from
the fact that

p(E)p(H|E) = p(E ∧ H) 6 p(H) = p(E ∧ H) + p(Ē ∧ H) 6 p(E ∧ H) + p(Ē),

along with easily constructed examples of probabilities p satisfying the antecedent
conditions of (7) and, respectively, (i) p(Ē ∧ H) = 0 and (ii) p(Ē ∧ H̄) = 0 Hail-
perin [H96] has also given an interesting proof using linear programming.

It follows from (7) that if 0 6 ε < 1, then

p(H|E) > 1 − ε and p(E) > 1 − ε ⇒ p(H) > (1 − ε)2, (8)

as noted by Suppes [S66].

3.3 Modus Tollens for Conditional Probabilities

Unlike the case of modus tollens for probabilities of conditionals, we cannot make
short work of the present subject by invoking the results of section 3.2. For it is not
generally true that p(E|H) = p(H̄|Ē). Indeed, from p(E|H) = a, with 0 6 a < 1,
one can conclude only that 0 6 p(H̄|Ē) 6 1, as shown by Hailperin [H96]. The
following theorem, which seems not to have appeared in the literature, delineates
what might reasonably be termed modus tollens for conditional probabilities.

Theorem 1 Let p(E|H) = a and p(Ē) = b.
If (i) 0 < a, b < 1, then

max{(1 − a − b)/(1 − a), (a + b − 1)/a} 6 p(H̄) < 1. (9)

If (ii) a = 0 and 0 < b 6 1, then

1 − b 6 p(H̄) < 1, and (10)

If (iii) a = 1 and 0 6 b < 1, then

b 6 p(H̄) < 1, (11)

and these bounds are the best possible.



Carl G. Wagner: Modus Tollens Probabilized 5

Proof. Let A be the algebra of propositions generated by E and H. Under each
of the conditions (i), (ii), (iii) above (and only under these conditions) the set P
of probabilities p on A satisfying p(E|H) = a and p(Ē) = b is nonempty and
consists of precisely those p such that p(E ∧ H) = at, p(E ∧ H̄) = 1 − b − at,
p(Ē ∧H) = (1−a)t, and p(Ē ∧H̄) = b−(1−a)t, where the parameter t is subject
to restrictions ensuring that these quantities are all nonnegative.

Under condition (i), we have

0 < t 6 min{b/(1 − a), (1 − b)/a}. (12)

Since p(H̄) = 1 −t, the inequalities in (9) follow from those in (12).
It is clear that the upper bound in (9) is the best possible. To show that this is

also true for the lower bound requires consideration of two cases. (1). If a + b 6 1
then b/(1 − a) 6 (1 − b)/a, and so t = b/(1 − a) is an allowable value of the
parameter t. For the probability p defined by this value of t, we have p(H̄) =
(1−a−b)/(1−a). Since a + b 6 1, we have (1−a−b)/(1−a) > (a + b−1)/a,
and so p(H̄) attains the lower bound in (9). (2). If a + b > 1, then (1 − b)/a <
b/(1 − a), and so t = (1 − b)/a is an allowable value of t. For the probability p
defined by this value of t, we have p(H̄) = (a + b−1)/a. Since a + b > 1, we have
(a + b − 1)/a > (1 − a − b)/(1 − a), and so p(H̄) again attains the lower bound
in (9).

Under condition (ii), we have

0 < t 6 b, (13)

which, with p(H̄) = 1 − t, implies (10), the lower bound in (10) being attained
when t = b.

Under condition (iii), we have

0 < t 6 1 − b, (14)

which, with p(H̄) = 1 − t, implies (11), the lower bound in (11) being attained
when t = 1 − b. 2

The above analysis reveals an interesting feature of modus tollens for con-
ditional probabilities, namely, that p(H̄) may be made as close to 1 as we wish
not only by making p(E|H) and p(Ē) sufficiently large, but also by making them
sufficiently small. This is of course intuitively reasonable. To take a very concrete
example, suppose that we have a finite set of spheres, each either hollow (H) or
solid (H̄), and each painted either ecru (E) or white (Ē). If most of the hollow
spheres are ecru, but most of the spheres are white, then most of the spheres must
be solid. Similarly, if few of the hollow spheres are ecru and few of the spheres
are white, then most of the spheres must be solid.

We conclude with an elaboration of the above in the spirit of Suppes’ inequal-
ities (4), (6), and (8).



6 Proceedings in Informatics

Corollary 1 Let 0 < ε 6 1/2. If

p(E|H) > 1 − ε and p(Ē) > 1 − ε, or (15)
p(E|H) 6 ε and p(Ē) 6 ε, then (16)

p(H̄) > (1 − 2ε)/(1 − ε). (17)

Proof. That (15) implies (17) follows from setting a = b = 1 − ε in (9) and
verifying by partial differentiation that (a + b − 1)/a is increasing in both a and
b. That (16) implies (17) follows from setting a = b = ε in (9) and verifying by
partial differentiation that (1 − a − b)/(1 − a) is decreasing in a and b. 2

Note that in each of the probabilistic inferences (4), (6), and (8) there is
some degradation in probability in passing from premisses to conclusion. The
same is true for the inference from (15) to (17), but this case of modus tol-
lens for conditional probabilities always involves less degradation in probabili-
ties than modus ponens or modus tollens for probabilities of conditionals and, if
ε < (3 −

√
5)/2

.
= 0.38, it involves less degradation than modus ponens for con-

ditional probabilities.

References

[A65] E. Adams, The logic of conditionals, Inquiry 8: 166–197, 1965.

[A75] E. Adams, The Logic of Conditionals, Reidel, Dordrecht, 1975.

[B54] G. Boole, An Investigation of the Laws of Thought, on Which are
Founded the Mathematical Theories of Logic and Probabilities, Walton
and Maberly, London, 1854.

[F35] M. Fréchet, Généralizations du théoreme des probabilités totales, Funda-
menta Mathematicae 25: 379–387, 1935.

[H65] T. Hailperin, Best possible inequalities for the probability of a logical
function of events, American Mathematical Monthly 72: 343–359, 1965.

[H84] T. Hailperin, Probability logic, Notre Dame Journal of Formal Logic 25:
198–212, 1984.

[H96] T. Hailperin, Sentential Probability Logic, Lehigh University Press, Beth-
lehem, 1996.

[J79] F. Jackson, On assertion and indicative conditionals, The Philosophical
Review 88: 565–589, 1979.

[L76] D. Lewis, Probabilities of conditionals and conditional probabilities, The
Philosophical Review 85: 297–315, 1976.



Carl G. Wagner: Modus Tollens Probabilized 7

[L86] D. Lewis, Probabilities of conditionals and conditional probabilities II,
The Philosophical Review 95: 581–589, 1986.

[S66] P. Suppes, Probabilistic inference and the concept of total evidence, in
Aspects of Inductive Logic, J. Hintikka and P. Suppes, eds, North-Holland,
Amsterdam, 49–55, 1966.

Carl G. Wagner is with the Department of Mathematics, University of Tennessee,
Knoxville, TN, U.S.A. 37996-1300. Email: wagner@math.utk.edu