CASPER J. ALBERS, BARTELD P.

KOOI and WILLEM SCHAAFSMA

TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM

ABSTRACT. After explaining the well-known two-envelope paradox by indicating the

fallacy involved, we consider the two-envelope problem of evaluating the factual in-
formation provided to us in the form of the value contained by the envelope chosen first.
We try to provide a synthesis of contributions from economy, psychology, logic, proba-
bility theory (in the form of Bayesian statistics), mathematical statistics (in the form of a
decision-theoretic approach) and game theory. We conclude that the two-envelope problem
does not allow a satisfactory solution. An interpretation is made for statistical science at
large.

1. INTRODUCTION

In 1943, Kraitchik discussed the paradox of the neckties:

Each of two persons claims to have the finer necktie. They call in a third person who must
make a decision. The winner must give his necktie to the loser as consolation. Each of the
contestants reasons as follows: I know what my tie is worth. I may lose it, but I may also
win a better one, so the game is to my advantage. How can the game be to the advantage
of both?

The snake in the grass is that equal probabilities or chances are assigned
to winning and losing. Kraitchik writes: In reality, however, the probabil-
ity is not an objectively given fact, but depends on ones knowledge of
the circumstances. In the present case it is wise not to try to estimate the
probability.
A similar paradox is the two-envelope paradox. It is unclear who gave
the problem its modern form; Kraitchik already formulated a problem
where two people compare the number of pennies in their purses. Gardner
(1982) called this the wallet game. Zabell (1988a, b) heard it from Budrys
(see Nalebuff 1989). Our formulation is a bit technical, this is useful later
on.
Two-envelope paradox. Two indistinguishable envelopes, 1 and 2, are
submitted to you (the decision maker; later the designation we is used
as the combination of you and the statistician who tries to assist you).
Envelope 2 contains a check worth twice the unknown value, say y, in
Envelope 1. You choose one of the envelopes, say Envelope z, at random

Synthese (2005) 145: 89109 Springer 2005

DOI 10.1007/s11229-004-7665-5
90 CASPER J. ALBERS ET AL.

and, after opening it, you observe the value x = zy contained by it. Finally
you are allowed to decide between
a1: keep the envelope you have.
a2: return this envelope and take the other one.
Two variants will be discussed. In Variant 1, the discrete case, you are told
that y N; Variant 2, the continuous case, tells you y R+ . A paradox
appears if you would argue that choosing at random implies that, as the
other envelope contains either 12 x (namely if z = 2) or 2x (when z = 1),
the expectation of the value contained by it is 12 12 x + 12 2x = 54 x, and that,
hence, swapping is advantageous on the average. The snake in the grass
is that you should not use the marginal or prior probability P(Z = 1),
but the conditional or posterior probability P(Z = 1|X = x), given the
knowledge you have, namely that X = Z Y has the outcome x observed.1
Unfortunately, the relevant posterior probabilities P(Z = z|X = x) (z =
1, 2) are unknown to you. Moreover you might question whether y can
really be regarded as the outcome of a random variable Y . The paradox
has been explained, but you are still in need of a solution to the problem.
Should you decide upon a1 or upon a2 , given the outcome X = x? That is
the question.
This problem has been discussed extensively in the literature, see, e.g.,
Zabell (1988a, b), Nalebuff (1989), and the (probabilistic) references cited
in Section 3. The literature shows that there are many ways to solve
this problem but, in the absence of additional information, these solutions
cannot be regarded as satisfactory.
Remark. One could suggest that the problem becomes more interesting
from the philosophical point of view if we discuss the contents of the
envelope chosen first without opening it. It is, of course, very natural for
the philosophical mind to keep thinking and to avoid the confrontation with
actual data. Note, however, that our formulation of the two-envelope prob-
lem is in some ways deeper and of more epistemological interest than that
where the envelopes remain unopened. From the perspective of our some-
what technical formulation it is clear that in the case of non-availability
of x, the paradox has its origin in the same fallacious assignment of
probabilities as before.
Note that choosing a1 leads to an expected amount E X = E (Z Y ) =
E Z E Y = 32 E Y which coincides with the amount E (3 Z )Y to be
expected from a2 . A fallacy appears if the latter amount is computed by
conditioning with respect to X and the relevant conditional probability
P(Z = 1|X ) is identified with the marginal probability P(Z = 1) = 12 .
This is erroneous because X = Y Z and Z are not statistically independent.
 TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM 91

2. EXPLORATIONS

It is obvious that you are in need of some additional information, fac-

tual, contextual, or otherwise. To settle the issue scientifically you might
consult a variety of experts. As money is involved, you might start with
an economist. He will argue that the utility of an action is not necessar-
ily proportional to its monetary value. The law of diminishing returns
prescribes this relation to be a concave function. However, if you have
observed x = 4.50, are hungry, and know that the cheapest pizza costs
7, then you will swap. Henceforth, it is assumed that there is complete
correspondence between utility and monetary value. Next the psychologist
might be consulted. He will tell you that some people are risk-averse (they
tend to prefer a1 ), others are risk-prone (they prefer a2 ). To proceed in the
most general way we assume that the decision maker has a risk-neutral
attitude and, hence, will try to maximize expected utilities (see Section
6 for a more elaborate treatment of the economist, psychologist, etc.).
This, however, requires probabilistic terminology. The logician Smullyan
(1997) maintained that probability is really quite inessential to the heart
of the two-envelope paradox. He presents it as a logical paradox, i.e.,
two contrary, or even contradictory, propositions to which we are led by
apparently sound arguments. The arguments are considered sound because,
when used in other contexts, they do not seem to create any difficulty
(Heijenoort 1967). Here are the propositions derived by Smullyan:

Proposition 1: The amount you will gain by trading, if you do gain, is

greater than the amount you will lose, if you do lose.

Proposition 2: The two amounts are really the same.

He proves both of them in the following way:2 To prove Proposition 1,

let x be the amount you are now holding. Then the other envelope either
contains 2x or x/2. If you gain by trading, you will gain x dollars (moving
from x to 2x), whereas if you lose by trading, you will lose only x/2. Since
x is greater than x/2, then Proposition 1 is established.
To prove Proposition 2, let d be the difference between the two amounts
in the envelopes (or what is the same thing, the lesser of the two amounts).
Well, if you gain on the trade, you will gain d dollars. If you lose on the
trade, you will lose d dollars. Since d is equal to d, then Proposition 2 is
established (Smullyan 1977, p. 174).
Both proofs seem to be sound, but it cannot be the case that both are
correct. The problem turns out to be that some of the terms that are used in
both proofs are ambiguous, they can be interpreted in two different ways.
92 CASPER J. ALBERS ET AL.

If you read the proofs you tend to interpret these terms in each proof in
such a way that the proof is correct. But if you take the interpretation that
makes one of the proofs correct and use that interpretation in the other
proof than the other proof is not correct.
The ambiguous terms are the amount you will gain by trading, if you
do gain and the amount you will lose by trading if you do lose. What
do these terms refer to? This all depends on what you mean by if you
do gain for example. Under what circumstances do you gain? In the first
proof this is the case if the amount in the envelope you did not pick is
double the amount you are now actually holding. In the second proof this
is the case if you picked the envelope which contains the highest amount
actually available.
The difference between these interpretation becomes clear if we look
at the case where you lose by trading. Now in the case of proof one and
the first interpretation, the amount you gain by trading, if you do gain
refers to twice the amount that is actually in your envelope, which is four
times what is in the other envelope. In case of proof two, on the other hand,
the amount you gain by trading, if you do gain refers to the amount you
would win by trading if you had picked the other envelope, which is one
times what is in the other envelope. A similar analysis is provided in Chase
(2002).
Regardless of which interpretation you choose, this does not resolve
the problem.
To make the choice between a1 and a2 additional knowledge will be
needed. In this respect it is natural to refer, as we shall do in Section 3,
to the probabilistic knowledge that P(Z = 1) = P(Z = 2) = 12 . It will
turn out that this factual knowledge does not yet settle the issue. That is
why we will look for more additional knowledge, perhaps of a less factual
kind, in the hope that, after all, some reasonable solution will appear. It
is in this respect that we will refer to the following logical paradox. The
point we shall make is that the context of this paradox enforces a solution
(at least temporarily).

The Protagoras paradox (Gellius 1946)

Euathlus, a wealthy young man, was desirous of instruction in oratory and the pleading of
causes. He became a pupil of Protagoras, the keenest of all sophists, and promised to pay
him a large sum of money, as much as Protagoras had demanded. He paid him half of the
amount at once, before beginning his lessons, and agreed to pay the remaining half on the
day when he first pleaded before jurors and won his case. Afterwards, when he had been
for some little time a pupil and follower of Protagoras, and had in fact made considerable
progress in the study of oratory, he nevertheless did not undertake any cases. And when the
time was already getting long, and he seemed to be acting thus in order not to pay the rest
 TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM 93

of the fee, Protagoras formed what seemed to him at the time a wily scheme; he determined
to demand his pay according to the contract, and brought suit against Euathlus.
And when they had appeared before the jurors to bring forward and to contest the case,
Protagoras began as follows: Let me tell you, most foolish of youths, that in either event
you will have to pay what I am demanding, whether judgment be pronounced for or against
you. For if the case goes against you, the money will be due me in accordance with the
verdict, because I have won; but if the decision be in your favour, the money will be due
me according to our contract, since you will have won a case.
To this Euathlus replied: I might have met this sophism of yours, tricky as it is, by not
pleading my own cause but employing another as my advocate. But I take greater satis-
faction in a victory in which I defeat you, not only in the suit, but also in this argument
of yours. So let me tell you in turn, wisest of masters, that in either event I shall not have
to pay what you demand, whether judgment be pronounced for or against me. For if the
jurors decide in my favour, according to their verdict nothing will be due you, because I
have won; but if they give judgment against me, by the terms of our contract I shall owe
you nothing, because I have not won a casex.

This is the paradox. But the context is such that Protagoras went to court.
This had the following consequences according to Gellius:

Then the jurors, thinking that the plea on both sides was uncertain and insoluble, for fear
that their decision, for whichever side it was rendered, might annul itself, left the matter
undecided and postponed the case to a distant day. Thus a celebrated master of oratory
was refuted by his youthful pupil with his own argument, and his cleverly devised sophism
failed.

We conclude that, due to the context within which it appeared, this paradox
is solved, from a practical point of view, in favor of Euathlus. Later authors,
e.g., Stewart (2000), missed this point. Dealing with the two-envelope
paradox we shall look for similar sources of additional information, us-
ing probability theory (Section 3), mathematical statistics (Section 4), and
game theory (Section 5).
Remark. One may not be satisfied by our presentation of the Protagoras
paradox which, indeed, is nothing but a translation of something archaic.
In consequence, one may see no reason to put faith in the moral we draw
from the paradox and which is that as the purely rational approaches of
Section 3 and part of Section 4 fail to be applicable, we may still try to
obtain something practical. With respect to the two-envelope problem
we completely agree that there is not much reason to put faith in this
moral because the factual information (the outcome x and the knowledge
that P(Z = 1) = 12 ) is very weak. In the beginning of Section 4 some
reference will be made to a situation where relevant additional information
is available. In Section 5 other additional information, only partly factual,
will be incorporated.
94 CASPER J. ALBERS ET AL.

We shall not analyze the Protagoras paradox here more carefully, log-
ically and mathematically: the mere reason of mentioning this paradox
was for making the statement that sometimes a temporary solution exists
suffices.

3. PROBABILITY THEORY

Apart from Smullyans, most attempts to solve the two-envelope prob-

lem are in a Bayesian spirit (e.g., Zabell 1988a, b; Nalebuff 1988, 1989;
Broome 1995; Christiansen and Utts 1992; Clark and Shackel 2000; Jack-
son et al. 1994; Linzer 1994; and McGrew et al. 1997). This is just a
short list from the large and expanding number of Bayesian papers written
on this subject. The formulation of the two-envelope problem chosen in
Section 1 is such that the gain or utility

U (x, y, z; a1 ) = x = yz

2x if z = 1
U (x, y, z; a2 ) = 2 32z
x = y(3 z) = 1
2
x if z = 2

depends on the true values x, y, and z governing the actual experiment

(one of these three can be deleted because x = yz). To incorporate the
information that P(Z = 1) = P(Z = 2) = 12 , we regard (x, z) as the
outcome of a pair of random variables (X, Z ) which, in principle, may
assume any value (, ) X {1, 2} where X is N or R+ . Using the
utility function in the form


if a = a1
U (, ; a) =
232 if a = a2

it is rational to choose a such that the conditional (or posterior) expecta-

tion


x if a = a1
E (U (x, Z ; a)|X = x) =
2xP(Z = 1|x) + 2 xP(Z = 2|x) if a = a2
1

is maximum or, equivalently, to use the procedure d : X {a1 , a2 }

defined by


a1 if P(Z = 1 | X = ) < 13
d ( ) =
a2 if P(Z = 1 | X = ) > 13 ,

the choice in the case of equality being arbitrary. Procedure d is such

that the expected utility is maximum, both a priori (unconditionally) and a
 TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM 95

posteriori (conditionally, given X ). Unfortunately d is not yet a workable

procedure: the conditional probability P(Z = 1|X = x) has not yet been
specified. Such specification requires additional information, e.g., about
the way y has come into being. Can y be regarded as the outcome of a
random variable Y and, if so, can the distribution of Y be specified? The
personalist Bayesian will answer both questions affirmatively and argues
as follows:
In the discrete case, let f () = P(Y = ) be specified ( Y). The
joint distribution of (X, Y, Z ) is then determined by (1) X = Y Z , (2) Y
and Z are stochastically independent, (3) Y has density f , (4) P(Z = 1) =
P(Z = 2) = 12 . We obtain

P(Z = 1, X = x)
P(Z = 1|X = x) =
P(X = x)
P(Z = 1, Y = x)
=
P(Z = 1, Y = x) + P(Z = 2, Y = 12 x)
f (x)
= .
f (x) + f ( 12 x)

If x is odd, then f ( 12 x) = 0 and {Z = 1} is sure: swapping provides you

with 2x. If x is even, then it depends on f whether (or not) f ( 12 x)/ f (x)
is smaller (or larger) than 2 and swapping (or not swapping) is most
profitable.
In the continuous case the difference between odd and even disappears
and we have

P(Z = 1|X = x) = lim P(Z = 1|x  < X < x + )

0
P(|x Y | < , Z = 1)
= lim 2
z=1 P(|x zY | < , Z = z)
0
f (x)
= ,
f (x) + 12 f ( 12 x)

(see Broome 1995). Now the optimal procedure prescribes to decide upon
a2 if f ( 12 x)/ f (x) is smaller than 4.
The Achilles heel of this (personalist) Bayesian solution is the assump-
tion that y is the outcome of a random variable Y with known density f .
Concerning f one can distinguish three cases: (i) that where f is fully
known (and above mathematics are directly applicable), (ii) that where
f is not completely known, but not complete unknown either (and some
other solution has to be found), and (iii) that where it is not reasonable to
96 CASPER J. ALBERS ET AL.

specify f (and, thus, no inference can be given). In Section 4 we shall con-

tinue this discussion (see also Section 6). In the remainder of this section,
a matter of theoretical interest is discussed.
Suppose that f is such that, in the discrete case f ( 12 x)/ f (x) 2 holds
for all even values of x, or that, in the continuous case, f ( 12 x)/ f (x) 4
holds for all x. Swapping is then always indicated according to the theory.
This sounds paradoxical because if you get Envelope z and I get Enve-
lope 3 z, we will both believe to gain by swapping. The explanation
is that this situation can only appear if E Y = and we both may
expect an infinite amount of money, no matter what we do. The existence
of paradoxical f s is a matter of elementary analysis. In the discrete case,


(h 1) h (s1) h > 0 if = 2s , (s = 0, 1, 2, . . .)
f () =
0 otherwise

works. Nalebuff (1989, p. 189) introduced this density (with h = 3/2),

and Broome (1995) linked this density to Daniel Bernoullis St. Petersburg
Paradox. In the continuous case


0 if < 0
f () = h
(h 1)( + 1) , h (1, 2], if 0

will do, the case h = 2 was mentioned by Broome (1995).

4. MATHEMATICAL STATISTICS

The previous section is based on the assumption that y is the outcome of

a random variable Y and that the distribution of Y is specified by a known
function f . In practice, this assumption is extremely questionable. It might
happen, of course, that previous experiences provided a reliable estimate
of f . Suppose, for example, that additional information is available in
the form of past outcomes x1 , x2 , . . . , xn of the content of the envelope
inspected first, in similar games. One may then try to make a reasonable
assessment of f . In such cases, the Bayesian solution may be interesting.
The thrill of the two-envelope problem, however, is in the situation that the
decision-maker has no other information than that the outcome of X is x
and, of course, that P(Z = 1) = P(Z = 2) = 12 . Some Bayesians will,
nevertheless, specify a distribution f , e.g., Linzer (1994) who concludes
his article as follows: So in one sense the answer may be that youve
got to guess on the probability distribution that the host is using and de-
cide accordingly. Of course, pure guessing is an unscientifically way to
 TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM 97

proceed. That is why, in the absence of sufficiently accurate additional

information, the Bayesian solution provides no real solution. Theoreticians
are fascinated by the seemingly simple problems of complete absence of
additional information. They will try to enforce a solution by using their
mind. In the present section we start out with two rationalizations, both
providing a unique solution (though these solutions should not be re-
garded as satisfactory). At the end we establish that one of these solutions
(namely swap, no matter x) is inadmissible from the viewpoint of Walds
theory of statistical decision functions.
Approach 1. At the beginning of Section 3 it was established that the
optimal procedure prescribes to assign a1 (a2 ) if the posterior prob-
ability P(Z = 1|X = x) is smaller (larger) than 13 . The question is
how to estimate this posterior probability. Note that, in Kolmogorovs
theory, this conditional probability is interpreted as the conditional expec-
tation E (1{Z =1} |X = x) and that, in general, the conditional expectation
E (1{Z =1} |X ) is the projection of 1{Z =1} on the space of all functions of X ,
a subspace of L 2 (, F , P ). This suggests to estimate P(Z = 1|X = x)
by constructing a procedure d : X [0, 1], this construction being such
 2
that the mean squared error of prediction E 1{Z =1} d(X ) is minimum.
The value d(x) is then the estimate required.
The joint distribution of X and Z is determined by that of (X, Y, Z )
discussed in Section 3. The density f of Y appears as the unknown para-
meter governing the risk. In the continuous case, absence of information
about f allows an interpretation in the sense of invariance under scale
transformations. If the predictor d is required to be scale invariant then it
is constant and the solution d 12 appears as uniformly, best invariant
predictor of 1{Z =1} . As 12 > 13 , its consequence is to swap, no matter
the outcome x observed. The arguments involved are considerably ma-
nipulative because the real issue, namely that of deciding whether or not
P(Z = 1|X = x) is smaller than 13 , is replaced by another one, namely the
prediction of 1{Z =1} . Moreover some information about f will exist.
Approach 2. All Bayesian and almost all non-Bayesian statisticians
know that invariance considerations are elegant as well as dangerous. They
would prefer to incorporate additional information, e.g., that E Y is ap-
proximately equal to some a priori value, say E Y = 100, or, if such value
is unavailable, an a posteriori value, say E Y = 23 x (because E X = 32 y).
Such information E Y = ( specified) can be incorporated elegantly by
constructing f such that the entropy
 
y=1 f (y) log f (y) (discrete case)
I( f ) = 
0 f (y) log f (y) dy (continuous case)
98 CASPER J. ALBERS ET AL.

is maximum under the restriction E Y = . It is well known that the

solution to this optimization problem is of the exponential form f (y) =
exp( y ()) where


() = log( exp( y)) = log(1 e )
y=1

in the discrete case, and


() = log( exp( y) dy) = log()
0

in the continuous case, and (, 0) is determined such that E Y =

() = . The interesting case is that where, in absence of further in-
formation, = 23 x is used and the ratio f ( 12 x)/ f (x) = exp( 12 x) is
compared with the value 2 in the discrete case when x is even (if x is odd,
then swapping is always indicated), and with the value 4 in the continuous
case.
In the discrete case we have that () = (1 e )1 is equal to 23 x if
1
= log(1 2x3 ) and f ( 12 x)/ f (x) = (1 2x3 ) 2 x is larger than 2 for all
x {2, 4, 6, . . .} and, hence, swapping if x is odd and not swapping if x is
even is indicated.
In the continuous case we have that () = 1 = 23 x implies =
2x3 and that f ( 12 x)/ f (x) = exp( 34 ) is smaller than 4 for all x and, hence,
swapping is always indicated.
Inadmissibility considerations. The above mathematical-statistical dis-
cussions were considerably manipulative. Two procedures were suggested
namely


a1 if is even
d( ) =
a2 if is odd

in the discrete case (see Section 5 for an extensive discussion), and

d( ) a2

in the continuous case. It is easy to establish that the latter procedure,

as well as that where d( ) a1 , is inadmissible in the sense that other
procedures exist with expected utility never smaller and often larger than
the expected utility 32 E Y of these constant procedures. An example is


a1 if > k
dk ( ) =
a2 if k
 TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM 99

with k a predetermined constant in R+ . To establish that dk and many

other procedures provide expected utilities above 32 E Y , it is of inter-
est to consider randomized procedures or, equivalently, test functions
: X [0, 1] with the interpretation that (x) is the probability of
a1 and (x) = 1 (x) that in favor of a2 . The utility of such randomized
decision is

U (x, y, z; a1 )(x) + U (x, y, z; a2 )(x) = x(x) + (3y x)(x)

if the outcome of (X, Y, Z ) is (x = yz, y, z). The expected utility is

E (X (X ) + (3Y X )(X )) = E (2X 3Y )(X ) + E (3Y X )

= E (2X 3Y )(X ) + 32 E Y

and the excess over the expected utility 32 E Y based on d( ) a1 or

d( ) a2 is

E (2X 3Y )(X ) = 1
2
E (2Y 3Y )(Y ) + 12 E (4Y 3Y )(2Y )
= 1
2
E Y ((2Y ) (Y ))

and this is strictly positive if is strictly increasing.

If = 1(k,) corresponds to dk , then the excess is equal to
1
2
E Y (1(k,) (2Y ) 1(k,) (Y ) = 12 E Y 1(k/2,k) (Y ) which is strictly positive
if P( 12 k Y k) > 0.
Conclusion. We still are not in the situation that the problem can be
regarded as solved. In this respect it is interesting to recall that the context
of the Protagoras paradox mentioned in Section 2 enforced a (temporary)
solution. A contextual ingredient already available but not yet explored is
that illustrated by Kraitchiks question at the beginning of Section 1: how
can the game be to the advantage of both? The next section is concerned
with this perspective.

5. GAME THEORY

We restrict the attention to the formulation of the problem involving only

two players: you/we and the mysterious player who completes the en-
velopes. This situation is very similar to the ones studied by Wald (1964) in
his decision-theoretic approach to the problems of statistics. Our approach
is in line with Ferguson (1967), and will use the notation style customary
in the theory of statistical decision functions.
100 CASPER J. ALBERS ET AL.

If the reader has the idea that this game-theoretic perspective will not
be very helpful in the present context, then he is right, as we shall see in the
rest of this section. He is invited to continue with Section 6 where impor-
tant interpretations are made. (The decision-theoretic NeymanPearson
Wald approach is, of course, of considerable interest elsewhere; the point
is that a sufficient amount of additional information should be available to
have faith in any approach.)
A game being defined as a triple, the first game (A, Y, U ) to consider
is that where Player 1 chooses an action from A = {a1 , a2 }, Player 2 (the
rich eccentric or Nature) chooses y from Y (either N or R+ ) and


yz if a = a1
U (y, z; a) =
y(3 z) if a = a2
goes from Player 2 to Player 1 (you/we). This payoff depends on the
outcome z of Z . This formulation is inadequate in the sense that Player 2
cannot be regarded as the minimizing player and also in the sense that
the information x = yz available to Player 1 has not been used. To allow
for the last mentioned statistical input, the game (A, Y, U ) is replaced by
the game (D, Y, U ) where we now have that Player 1 chooses a decision
procedure d : X A from the class D of all (nonrandomized) rules of
this kind. The payoff is defined as
U (y; d) = E U (y, Z ; d(x, Z ))

y if d(y) = a1 ,d(2y) = a2
= 2y if d(y) = a2 ,d(2y) = a1
3
2
y if d(y) = d(2y)
This formulation still is irrelevant in the sense that Player 2 will choose
y as small as possible. If in the discrete case the specification Y =
{1, 2, . . . , n} is made (for some given n) then taking y = 1 is the min-
imax strategy of Player 2 and any rule with d(1) = a2 and d(2) = a1
is such that the minimum payoff is maximum. The two-person zero-sum
game (D, Y, U ) is determined in the sense that maxd min y U (y; d) =
min y maxd U (y; d) = 2 while the saddle points are characterized in the
above. Randomization can be dispensed with.
In his review of Wald (1964), Savage (1951) explained and criticized
the preoccupation by losses, risks, errors, etc., in the NeymanPearson
Wald approach to statistics. In this approach we are not discussing utilities
but regrets; shortcomings with respect to the best. Given y, the best we can
do is to swap if we observe y and to keep what we have if we observe 2y.
Hence
L(y, z; a) = max U (y, z; ah ) U (y, z; a)
h=1,2
 TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM 101

= 2y U (y, z; a)


(2 z)y if a = a1
=
(z 1)y if a = a2

and the expected loss, given y, is equal to

y if d(y) = a1 , d(2y) = a2
R(y; d) = 0 if d(y) = a2 , d(2y) = a1
1
2
y if d(y) = d(2y).

If the reader has the idea that this loss-function reformulation is too manip-
ulative to be satisfactory, then we agree. The reader, again, might continue
with Section 6. The analysis, however, is not uninteresting, as we shall see
below.
After this Umwertung aller Werte, the rich eccentric has become
Player 1 while we are now Player 2, the minimizing player in the two-
person zero-sum game where we are allowed to choose from the set D and
pay the amount

R(y, d) = E L(y, Z ; d(X ))

where X = y Z and this risk depends on the value y chosen by Player 1, the
equivalent of Nature in Walds approach to the problems of statistics, ini-
tiated nicely in his review of Neumann and Morgenstern (1944; see Wald
1947). The game (Y, D, R) is not determined because the lower value of
the game

v = sup inf R(y, d) = 0,

yY dD

is less than the upper value

v = inf sup R(y, d) > 0.

dD yY

In regular situations, it is a necessity that Y is finite for v < .

Exceptions (irregular situations) are special constructions such as Y =
{1, 3, 5, . . .} where, of course, the optimal solution


a1 , x even
d(x)
a2 , x odd

yields R(y, d) = 0 for all y.

102 CASPER J. ALBERS ET AL.

When no restrictions are put on Y then v = because

max(R( n2 , d), R(n, d)) = n
4
in the nonrandomized case, and
max(R( 2 , ), R(n, )) = 6 after randomization.3
n n

To overcome this difficulty and to establish a positive result we will

discuss the specific case where the upper bound n is specified a priori.
We will review both the continuous (Y = [0, n]) and the discrete case
(Y = {1, 2, . . . , n}).
The case Y = [0, n], n given. We are interested in constructing the min-
imax risk procedure d (if it exists) and start out by noting that Bayesian
and Laplacian statisticians are attracted by the uniform prior with density
f () = n 1 , (0 n). Section 3 provides that the corresponding
Bayes rule prescribes


a2 if x n
d (x) =
a1 if x > n

which, quite surprisingly, is minimax as well, with respect to the class D

of all nonrandomized Bayes rules. To establish this minimaxity of d , we
note that

1
y if 0 < y 12 n
R(y, d ) = 2
0 if y > 12 n

and that, hence, sup y[0,n] R(y, d ) = 14 n. Next we concentrate the atten-

tion on the two possibilities y = 12 n and y = n having x 12 n, n, 2n
as consequence. Any nonrandomized rule d assigns values d( 12 n), d(n),
d(2n) to these outcomes. The most appropriate assignments (a2 , a1 , a1 )
and (a2 , a2 , a1 ) lead to maximum risks equal to 12 n and 14 n, respectively.
Hence sup y R(y, d) 14 n holds for all d. Equality holds for d which,
hence, is minimax in the sense that

sup R(y, d ) = inf sup R(y, d)

yY dD yY

= v.

Note that minimaxity holds with respect the the class D of nonrandomized
mixed rules. If randomization is allowed, the upper value v of the game
can be decreased to n6 by taking
1
3 , x n2
(x) = 23 , n2 < x n

1, x > n
 TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM 103

The case Y = {1, . . . , n}, n given. Following the suggestions in Fer-

guson (1967), we consider the mixed extension (Y , D, r ). Here Y is the
class of all probability measures on Y, each one characterizable by an el-
ement f from the unit simplex Sn in Rn , the coordinates f corresponding
to the probabilities assigned to the possibilities for y. The class D is that
of behavioral randomized rules : X A which, as indicated before,
prescribe that an action is taken from A = {a1 , a2 } according to a random
mechanism which chooses a1 with probability

(x) = (x)({a1 }),

where the test function : X [0, 1] with

{1, 2, . . . , n, n + 2, n + 4, . . . , 2n} if n is even
X=
{1, 2, . . . , n, n + 1, n + 3, . . . , 2n} if n is odd.

The loss involved in such randomized decision is defined as

L(y, z, (x)) = (x)L(y, z; a1 ) + (1 (x))L(y, z, a2 )

= (x)(3y 2x) + x y

where, of course, x = yz. The risk (expected loss) given y is equal to

R(y, ) = E (y Z )(3y 2y Z ) + yE (Z 1)
= 12 y(1 + (y) (2y)).

with special cases (needed later on)

R( n2 , ) = 1n
22
(1 + ( n2 ) (n)) = n6
R(n, ) = 1
2
n(1 + (n) (2n)) = n6 ,

if n is even. In case n odd, we have R( n1 2

, ) = R(n 1, ) = n1
6
. Note
that there is no reason not to choose (2n) = 1.
To minimize max y{1,...,n} R(y, ) (and for some other purposes as well)
it is interesting to consider the Bayes risk


n
r (, ) = R(y, ) f (y)
y=1

= 1
2
f T (a + A)
104 CASPER J. ALBERS ET AL.

where a = (1, 2, . . . , n)T and A is equal to

1 1 0 0 0 0
  1 1 0 0 0
1 1 0 0 2 0 2 0 0
, 0 2 0 2 0 ,
0 0 3 0 3 0
0 2 2
0 0 3 0 3
0 0 0 4 0 4
in the cases n = 2, 3, 4. Extensions to n 5 are obvious.
Any of the general minimax risk theorems of Nikaid (1953, 1959; see
Schaafsma 1971 for more references) implies that
max min r ( f, ) = min max r ( f, ) =
f f

and that a saddle point ( f , ) exists such that

r ( f, ) = r ( f , ) r ( f , )
holds for all f Sn and . Here we used that r : Sn R defined
by
r ( f, ) = 1
2
f T (a + A)
is linear in f and affine-linear in . The theorems imply that the procedure
is minimax and the distribution with density f is least favorable.
The saddle point ( f , ) is not necessarily unique and in our problem
it is not unique (if n 2). We are now, fortunately, well equipped to
characterize the minimax risk procedures.
THEOREM 5.1. If n is even, then ( n2 ) = 0, (n) = 1
3
is necessary to
have max R(, ) n6 . If we choose such that

0 if 1 i n2
(i) = 31 if n2 < i n

1 if n < i 2n
accordingly, then we have max R(, ) = n
6
. The class of all minimax
risk procedures is characterized by
{ = (1 , 2 , . . . , n , n+2 , . . . , 2n
T
)| max 21 (1 + () (2)) n6 }.

If n is odd, then ( n1
2
) = 0, (n 1) = 1
3
is necessary to have
max R(, ) n16
. If we choose such that


0 if 1 i n1
1 2
if n1 <i n1
(i) = 3 2

0 if i = n

1 if n < i 2n
 TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM 105

then we have max R(, ) = n1

6
. The class of all minimax risk
procedures is characterized by
{ = (1 , 2 , . . . , n , n+1 , n+3 , . . . , 2n
T
)| max 21 (1 + () (2))

n1
6
}.

Proof. It follows from max(R( n2 , ), R(n, )) = n

6
that

minD max R(, ) 6 . If we choose = as mentioned, then
n

max R(, ) = n6 . It is easily seen that

R(, ) =1
(1 + () (2))

2
12 if n4
= 1
if n4 < n2
31
6
if n2 < n

The maximum risk will thus be n8 in case y n4 , and n6 in cases n4 < y n2

and n2 < y n. Hence, max R(, ) = n6 .
That this solution is not unique is quickly seen by taking equal to
, with the exception that (1) = 1. Then

1 if = 1
R(, ) =
R(, ) elsewhere

and (if n > 6) no harm is done to the maximum.

In case of n odd, then the necessity of ( n1
2
) = 0, (n 1) = 13 is
easily shown, in the same way as for the even case. That max R(, )
n1
6
is trivial after the observation that

R(, ) = 1
(1 + () (2))
1
2

if n1
12 4
if n1 < n1
= 3 4 2


1
if n1 <n 1
6 2
0 if =n

Analogue to the case n even, the solution and its non-uniqueness are trivial.

6. DISCUSSION : THE LIMITS OF REASON

Stripped to its logical essentials, the two-envelope problem is that of

choosing between the sure gain x and the unsure gain that is either 12 x
106 CASPER J. ALBERS ET AL.

or 2x. Issues like the one discussed in the two-envelope problem cannot be
settled unless something additional is incorporated. The Protagoras para-
dox is of particular interest in this respect because, there, the context was
such that a (temporary) solution was enforced. With respect to the two-
envelope problem one has to go beyond the logic essentials because the
knowledge that P(Z = 1) = P(Z = 2) = 12 has to be incorporated.
The Bayesian solutions obtained, however, are not applicable because they
depend on the unknown function f . The really interesting element of this
two-envelope problem is, however, that nothing is known about f (even not
if such an f actually exists). In Section 4 an attempt was made to settle the
issue mathematically. These attempts were not successful. In Section 5 we
tried to incorporate arguments from the theory of games and from Walds
theory of statistical decision functions. It was only after the specification of
the upper bound n for y, that we could arrive at something not completely
unreasonable.
We conclude that, if almost no information exists with respect to f , then
it is wise not to try to estimate the posterior probability P(Z = 1|X = x).
Note that already more than half a century ago, Kraitchik (1943) made
a similar statement, see Section 1. If an optimal decision or optimal pro-
cedure is advocated then a scrutiny will reveal that it largely is based on
something fictitious. It would be a fallacy of misplaced concreteness if one
accepts such result as sufficiently compelling or valid.
We are interested in the two-envelope problem, because there are some
consequences for Statistical Science at large. It often happens that the sta-
tistician is asked to use data in order to compute some posterior probability,
to make a distributional inference, or to suggest an optimal decision. Some,
perhaps many, of these situations are such that the lack of relevant infor-
mation is so large that it is wise not to try to settle the issue. This leaves
us with the problem to draw a distinction between those situations where
the information is too weak to say something and those where the informa-
tion is sufficiently overwhelming. The difficulty is, of course, in the area
between. (See also the beginning of Section 4 where reasonable Bayesian
or non-Bayesian solutions will emerge if the value of x is available for a
sufficiently large number of similar games.)
The criticism mentioned here does not only refer to the Bayesian ap-
proach where it has to be considered unwise to specify a prior distribution
unless relevant information is sufficiently abundant. Criticism refers also to
the Neyman-Pearson-Wald approach where one is trying to discuss many
possible worlds, one for every value of the parameter . Usually we as-
sume that exactly one of these worlds is factually true, but we dont know
which one. We, perhaps, have to admit that our approaches are consid-
 TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM 107

erably manipulative and, as Karl Pearson noticed in his paper about the
Fundamental Problem of Practical Statistics (1920), in the nearness of
an abyss.
In our approach to the two-envelope problem, the economist, psy-
chologist and logician did not get as much attention as the probabilist,
mathematical statistician and game theorist, because we considered P(Z =
1) = P(Z = 2) = 12 to be knowledge which should not be ignored.
The two-envelope problem is an extremely simple example of the com-
plicated socio-economic and other issues to be settled in practice. Sciences
like economy, sociology, and psychology try to contribute to the settlement
of these issues by emphasizing various aspects of the managerial use of
information. The basis of everything is, of course, that data are collected
representing the state of reality. Unfortunately, such statistical data leave
room for uncertainties of various kinds. The sciences indicated have much
more to say than is suggested in the context of the two-envelope problem.
What they have to say, however, is largely of a qualitative existential
nature. The economist may infer from empirical data that assessments of
utilities and probabilities are considerably subjective, i.e., different from
person to person. The psychologist may use his data to characterize people
as systematically risk averse or risk prone. The sociologist will argue that
economic behaviour is much less rational than a decision-theorist might
hope. The two-envelope problem provides a nice illustration of the limits
of reason and the need of information to make inferences and decisions
reasonable.

ACKNOWLEDGEMENTS

The physicist L. K. ter Veld taught us about the Protagoras paradox. The
decision procedure E (2X 3Y )(X ) on page 99 arose during discus-
sions with W. Oudshoorn and R. Auer. The reviewers made very helpful
suggestions that led to improvements of the manuscript.

N OTES
1 We use standard probabilistic terminology; random variables are denoted by capital
letters and regarded as functions on some underlying probability space (, F , P ). Factual
outcomes are denoted by lowercase letters like x, y, z. If a priori possible outcomes are
discussed, then we use , , to denote the possibilities for x, y, z.
2 In his book Smullyan uses the letter n in his proof of proposition 1. We have replaced it
with x to keep a uniform notation.
3 See later in this section, is used as notation for a randomized rule.
108 CASPER J. ALBERS ET AL.

R EFERENCES

Broome, J.: 1995, The Two-Envelope Paradox, Analysis 55, 611.

Chase, J.: 2002, The Non-Probabilistic Two Envelope Paradox, Analysis 62, 157160.
Christiansen, R. and J. Utts: 1992, Bayesian Resolution of the Exchange Paradox , The
American Statistician 46, 274276.
Clark, M. and N. Shackel: 2000, The Two-Envelope Paradox, Mind 109, 415442.
Ferguson, T.: 1967, Mathematical Statistics, a Decision-Theoretic Approach, Academic
Press, New York.
Gardner, M.: 1982, Aha! Gotcha: Paradoxes to Puzzle and Delight, W.H. Freeman, New
York.
Gellius, A.: 1946, The Attic Nights of Aulus Gellius (Noctes Atticae), translated by J. C.
Rolfe, Harvard University Press, Cambridge, MA.
Heijenoort, J. v.: 1967, Logical Paradoxes in P. Edwards (ed.), The Encyclopedia of
Philosophy, Vol. 5, Macmillan Company & The Free Press, 4551.
Jackson, F., P. Menzies and G. Oppy: 1994, The Two Envelope Paradox , Analysis 54,
4345.
Kraitchik, M.: 1943, Mathematical Recreations, George Allen & Unwin, London.
Linzer, E.: 1994, The Two Envelope Paradox, The American Mathematical Monthly 101,
417419.
McGrew, T., D. Shier and H. Silverstein: 1997, The Two-Envelope Paradox Resolved,
Analysis 57, 2833.
Nalebuff, B.: 1988, Cider in Your Ear, Continuing Dilemma, the Last Shall Be the First,
and More, Journal of Economic Perspectives 2, 149156.
Nalebuff, B.: 1989, The Other Persons Envelope is Always Greener, Journal of
Economic Perspectives 3, 171181.
Neumann, J. and O. Morgenstern: 1944, Theory of Games and Economic Behaviour,
Princeton University Press, Princeton, NJ.
Nikaid, H.: 1953, On a Minimax Theorem and its Applications to Functional Analysis,
Journal of the Mathematical Society of Japan 5, 8694.
Nikaid, H.: 1959, On a Method of Proof for the Minimax Theorem, Proceedings of the
American Mathematical Association 10, 205212.
Pearson, K.: 1920, The Fundamental Problem of Practical Statistics, Biometrika 13, 116.
Savage, L.: 1951, The Theory of Statistical Decision, Journal of the American Statistical
Association 46, 5567.
Schaafsma, W.: 1971, The Neyman-Pearson Theory for Testing Statistical Hypotheses,
Statistica Neerlandica 25, 127.
Smullyan, R.: 1997, The Riddle of Scheherazade, and Other Amazing Puzzles, Ancient and
Modern, Knopf, New York.
Stewart, I.: 2000, Mathematical Recreations: Paradox Lost, Scientific American 6, 8889.
Wald, A.: 1947, Book Review of (Neumann and Morgenstern 1944), reprinted from The
Review of Economic Statistics 24, 4752.
Wald, A.: 1964, Statistical Decision Functions, John Wiley, New York.
Zabell, S.: 1988a, Loss and Gain: The Exchange Paradox in J. M. Bernardo, M. H.
DeGroot, D. V. Lindley and A. F. M. Smith (eds), Bayesian Statistics, 3. Proceedings
of the Third Valencia International Meeting, Clarendon Press, Oxford, pp. 233236.
Zabell, S.: 1988b, Symmetry and its Discontents in B. Skyrms and W. L. Harper (eds),
Causation, Chance and Crecedence, Kluwer Academic Publishers, Dordrecht.
 TRYING TO RESOLVE THE TWO-ENVELOPE PROBLEM 109

Mathematics & Natural Science

University of Groningen
P.O. Box 800
9700 AV Groningen
The Netherlands
E-mail: [email protected]