You are on page 1of 12

The Two Envelopes Paradox: Exploring Expected

Value and Rational Choice

David J. Runger
October 30, 2007

The feeling of surprise and discomfort that we experience when we en-
counter a paradox tells us that our understanding of the subject at hand is
somehow incomplete [2]. A paradox invites us—perhaps even shames us—
to deepen our understanding of the subject and to eliminate the apparent
contradiction that is the essence of any paradox. In the following pages, we
will grapple with the Two Envelopes Paradox, a paradox of probabilities.
One of the earliest presentations of this paradox was in 1988 as a puzzle
for readers of the Journal of Economic Perspectives. The readers’ response
was “overwhelming”, according to the editor of the “Puzzles” column, Barry
Nalebuff [5], and it’s easy to see why. Beguiling in its simplicity, the paradox
entices us with the prospect of a quick and easy resolution—but unraveling
the Two Envelopes Paradox is anything but child’s play.

The Paradox

Suppose that a game show host sets two envelopes before you, telling you
(truthfully, of course!) that each envelope contains a check written for a
positive, real-number sum and that one of the envelopes contains exactly
twice as much money as the other. Luckily for you, you are allowed to take
one of the envelopes home with you! What’s more, you can even look into
one (but only one) of the envelopes before you make your final decision about
which to take. Will you take the envelope into which you peek, or will you
switch your choice to the other envelope? Does it matter either way?


does 0 = 1? 2 . on average. if A contains the larger prize. So you can expect to win 25% more. since they must sum to 1. Thus. take a peek inside. If Envelope A contains the smaller prize. you might reason that you are just as likely to have chosen the greater prize as to have chosen the lesser prize. then B contains $a/2. Figure 1: To switch or not to switch? Let’s say that you choose Envelope A. Was that thinking somehow flawed? Is something wrong with our expected value calculation? Or. so you might as well stick with the envelope you have. we have E(B) = (1/2) · (2a) + (1/2) · (a/2) = 5a/4. Having seen how much money is in Envelope A. switching envelopes won’t improve your chances of finding the larger prize. thinking of B as a random variable representing the amount of money in Envelope B. undoubtedly the worst possible case. This reasoning certainly looks sound. then Envelope B contains $2a. you decide to double-check your conclusion by calculating the expected value of taking Envelope B instead. but having learned a lesson the hard way from a financially taxing trip to Las Vegas. If the probability that you chose the larger envelope is the same as the probability that you chose the smaller envelope. then the probability of each is 1/2. and find a check worth $a. should you switch your choice to Envelope B? Reflecting on the fact that you chose Envelope A randomly. just by switching! But this contradicts our original idea that switching envelopes would be useless.

even 3 . 5]. This assumption is indeed valid before you look in Envelope A and see how much money it contains. however. we must consider the underlying probability distri- butions for the amount of money in the two envelopes. Then. then you could conclude that the probability that Envelope B contains the larger sum is not 1/2. we can perform a correct expected value calculation [4. P (A = S) is the probability that Envelope A contains the smaller amount of money. given these distributions. Before you open Envelope A. the assumption is no longer necessarily true. but 0. and the chance that it contains the larger check is 1/2. so P (A = S) = P (A = L) = 1/2. and P (A = L) is the probability that Envelope A contains the larger amount of money. we must apply our knowledge of the amount of money in En- velope A to calculating the probability that Envelope B contains a/2 and the probability that B contains 2a. since it is impossible for B to contain $40 in a world in which only $30 exist.The Paradox Defanged and Dethroned The paradox can be resolved by noting an incorrect assumption made in the expected value calculation: the assumption that the probability that you chose the larger envelope is the same as the probability that you chose the smaller envelope. and neither necessarily equals 1/2. however. For example. suppose that the total amount of money in the world is $30. Specifically. having recalculated these proba- bilities with respect to our knowledge that A contains a. We must utilize the information gained by looking at the check in Envelope A. the paradox arises when we incorrectly use prior probabilities in our expected value calculation. so P (A = S|A = a) does not necessarily equal P (A = L|A = a). Now we are beginning to see the essential ideas required to resolve the paradox. the chance that it contains the smaller check is 1/2. (Thus. where S is a random variable representing the amount of money in the smaller envelope and L is a random variable representing the amount of money in the larger envelope. then. If you were to see $20 in Envelope A.) Once you see that the envelope contains $a. Essentially.

Thus. Rather. where only 4 . thereby resurrecting the paradox? (Of course. P (S > c/3) = 0 and P (L > 2c/3) = 0. it is not sufficient to say that P (A = S) = P (A = L) = 1/2 for all a. if the random variables S and L have upper bounds. the upper bounds on the ranges of S and L would be c/3 and 2c/3. for all a. S is a random variable whose range is the possible values of the smaller check. is not true for all a. then we know that P (A = S|A = a1 ) = 0 6= 1/2 and P (A = L|A = a1 ) = 1 6= 1/2. 2c/3]. If we find in Envelope A an amount a1 ∈ (c/3. on the total amount of money available to be put into the envelopes. Can S range from 0 to ∞? What if we suppose that the probability dis- tributions governing S and L are unbounded? Might it then be possible that. E(B) = 5a/4. That is. resolving the paradox (when we assume an upper bound on the amount of money that can be put in the envelopes). such a scenario is impossible in reality. where A is a random variable representing the amount of money in the en- velope we have chosen. it cannot be true that P (A = S|A = a) = P (A = L|A = a) = 1/2 for all a. c. for P (A = S|A = a) = P (A = L|A = a) = 1/2 for all a if we assume that there is an upper bound. respectively. consider that. P (A = S|A = a) = P (A = L|A = a) = 1/2. however. $a/2. To summarize this prose with a formula. and L is a random variable whose range is the possible values of the larger check [4]. When the ranges of S and L have upper bounds It is impossible. it must be true that P (A = S|A = a) = P (A = L|A = a) = 1/2 for all a. Thus. for E(B) to equal 5a/4 for all a.after we acquire relevant information that would allow us to determine more accurately the probability that Envelope B contains $2a vs. E(B) = P (A = S|A = a) · (2a) + P (A = L|A = a) · (a/2). the initial paradoxical expected value calculation. To see why. in this case. Therefore.

a finite amount of money exists. since P (A = S) = 1/2 and P (A = L) = 1/2. Furthermore. (2) Substituting (2) into (1).) Bayes’ Theorem tells us P (A = S|A = a) = P (A = a|A = S) · P (A = S)/P (A = a). Below is Keith Devlin’s proof that P (A = S|A = a) = P (A = S)—our necessary condition for rejuvenating our paradoxical expected value calculation—is true for all a only if the probability distributions governing S and L are uniform on the interval (0. So we must show that P (A = S|A = a) = 1/2 and P (A = L|A = a) = 1/2. ∞) [4]. this is not possible because this scenario would imply that S and L are uniform on the interval (0. no. we can see that P (A = a|A = S) = P (S = a). we must convert E(B) = P (A = S|A = a) · (2a) + P (A = L|A = a) · (a/2) into E(B) = (1/2) · (2a) + (1/2) · (a/2) = 5a/4. ∞) [4]. (1) We are able to apply Bayes’ Theorem in this case because the events A = S and A = L make up a disjoint partition of the sample space of possible ways that Envelope A could be characterized in our game—as either the envelope containing the smaller prize or that containing the larger prize.) The short answer is that. which is impossible [3]. we have P (A = S|A = a) = P (S = a) · P (A = S)/P (A = a). but the case is nonetheless interesting and informative to consider. (A proof that P (A = L|A = a) = P (A = L) implies the same conclusion proceeds in an almost identical fashion and could also be used [4]. (3) 5 . which is equivalent to proving that P (A = S|A = a) = P (A = S) and P (A = L|A = a) = P (A = L). In order for our original expected value calculation to be correct.

let p be a uniform probability R k+1 distribution on (0. To see this. This statement. however. the game 6 . we can see that P (A = S|A = a) = P (A = S)—our necessary condition for resurrecting the paradox—if and only if P (S = a) = P (A = a). in Envelope A. take P (S = a) = P (A = a) to be our new necessary condition. satisfies the R ∞ requirement of proper probability distributions that −∞ p(x)dx = 1. therefore. Since P (L = a) = P (S = a/2) and P (A = S) = P (A = L) = 1/2. again since A = S and A = L make up a disjoint partition of the sample space. P (S = a/2) = P (S = a). Recap We have seen that no conditions will allow us to calculate a correct expected value such that E(B) > a for all possible prizes. This means that k p(x)dx = c for all k and for some constant c. if c > 0. then the area under the curve is 0. It is impossible. for such a distribution to exist [3]. you shouldn’t always have an expected gain from switching to the other envelope. Neither of these. The paradox lay in the tension between two contra- dictory positions: 1. then the area under the curve is infinite. implies that the probability distribution of S must be uniform on (0. Since you select an envelope at random. If c = 0. ∞) [4]. (4) Thus. After all. in order to satisfy the necessary condition that P (S = a) = P (A = a). We can see that P (A = a) = P (A = S) · P (S = a) + P (A = L) · P (L = a). ∞). this becomes P (A = a) = 1/2 · P (S = a) + 1/2 · P (S = a/2).From (3). it must be true for every a that P (S = a/2) = P (S = a). a. We will now. thereby resolving the paradox. however.

A Strategy for Success We have shown that it is not always advantageous to switch envelopes. by switching envelopes every time. In the unbounded case. In the case of a bounded distribution. thereby resolving the paradox. the expected value of the prize in the other envelope is calculated to be E(B) = 5a/4. is symmetric. A. By scrutinizing the possible probability distributions that could define the likelihood that a given amount of money is in a given box. But when does switching envelopes have an expected gain? That is. on average. could we 7 . in the bounded case. we showed that item 2 is false. and therefore you should expect to increase your winnings. since the expected value calculation applies to any a we see. why should choosing Envelope A and then switching to Envelope B give us a better chance of success than simply choosing B in the first place? 2. it is easy to see that. Seeing the amount $a in your chosen envelope. for some values we might encounter when we look in Envelope A. ∞) in a way that meets the criteria of the problem. Further Puzzles Our dismantled paradox still invites a number of interesting questions. there is actually not a 1/2 chance that Envelope B contains the greater prize. Put another way. Therefore. we have seen that there can exist no proper underlying probability density function governing the random variable S (and also L) on (0. the expected value calculation fails for this reason. it is impossible even to perform a correct expected value calculation for Envelope B without having density functions that define the likely contents of B.

for L. 8 . In the case of a uniform probability distribution for L supported on (0. switching is profitable if and only if P (A = L|A = a) < 2/3. 0 Thus. P (A = S|A = a) = 1 − P (A = L|A = a)1 (6) Substituting (6) into (5) gives 2 · [1 − P (A = L|A = a)] + (1/2) · P (A = L|A = a) > 1. 2c/3). the amount of money we see when we peek into our initially selected envelope? This question is answered by Brams and Kilgour [1]. we can see that if we know the probability distribution function. and that A = S if and only if A 6= L. From the General Exchange Condition. p(x). or (3/2)P (A = L|A = a) < 1. Thus. Switching from En- velope A to B will increase our expected winnings if and only if the expected value of Envelope B is greater than a: E(B|A = a) = (2a) · P (A = S|A = a) + (a/2) · P (A = L|A = a) > a. given the probability distribution function of S (or L) and a. then we should switch if a < a0 and keep Envelope A if a > a0 . a0 will be 4c/9. Brams and Kilgour call this inequality the General Exchange Condi- tion [1]. if we see a in Envelope A. Thus.develop a strategy that tells us whether or not to switch envelopes. or 2 · P (A = S|A = a) + (1/2) · P (A = L|A = a) > 1 (5) We know that since A = S and A = L make up a disjoint partition. either A = S or A = L. then we can find an a0 which satisfies the equation Z a0 2/3 = p(x)dx = 1.

What would happen if we were to pit two rational players against each other. respectively. though. which is less than a. 2). and letting each look at the sum in his own envelope. I am sure to have the larger envelope (since the smaller envelope contains at most c).” 9 . though perhaps not paradoxical. The probability distributions governing S and L. A solution to this question.Make It Competitive We have so far presented the game as a one-player attempt to maximize profit in a game show. Let’s consider how Anne might reason: “I see a in my envelope. As we have seen. If a is between c and 2c. Would it ever happen that both players want to trade with each other? If so. 2). so switching would be foolish. which we’ll call 3c. c < a ≤ 2c implies that E(B) = a/2. We tell our two players. must be bounded by the constants c and 2c. Anne and Bob. is given below. giving one Envelope A and one Envelope B. I know that for any given trade with Bob. then. Put another way. in which case I might actually be able to double my money (indicated by the green mapping in Fig. I could conceivably either halve or double my money. the maximum amount of money that might be in the envelopes and give Envelope A to Anne and Envelope B to Bob. I won’t switch when a is between c and 2c—but I will offer to trade with Bob if a ≤ c. then this would appear to be a somewhat strange situation. so I would have no chance of doubling my money by trading with Bob. the total possible amount of money in the envelopes by a constant. Therefore. conceived by Edward Norton and presented in Nalebuff [5]. I would be guaranteed to halve my winnings (indicated by the red mapping in Fig.

When a > c/4. but I’m cleverer still. Just like Anne. in which case Anne might be willing to make a trade that is actually profitable to me. That’s pretty smart.” 10 . but it is possible for her to double her money when a ≤ c. So I will only switch if b ≤ c/2.Figure 2: Anne can’t profit from a trade when c < a ≤ 2c. Bob: “Anne must be thinking that she won’t trade if c < a ≤ 2c. After all. I won’t switch if c < b ≤ 2c—but I also won’t trade if c/2 < b ≤ c. Since the most that Bob will ever trade away is c/2. in order for me to profit from a trade when c/2 < b ≤ c. I will only have a positive expected value from trading when a < c/4. Anne: “I know what Bob is thinking: he’ll only switch if b ≤ c/2. in which case a trade to which Bob would acquiesce might double my money. Anne would have to be willing to trade me between c and 2c—but I know that she won’t. I know that I can’t possibly double my money. Figure 3: Bob anticipates that he will lose money if he trades when c/2 < b ≤ c.

as with the famous example of Russell’s 11 .” Figure 5: A function can be mapped from A ∪ B to Nn+m when A ∩ B = ∅ We can see that the reasoning continues inductively in this manner until neither player is ever willing to trade [5]. So I’ll have to limit my trading to the cases where b < c/8 in order to have a positive expected value from trading. the very mathematical systems within which we are operating are themselves flawed. in order to see where theory has been misapplied.Figure 4: Anne now thinks that she can only profit from trading when a < c/4.. as is the case with the Two Envelopes Paradox. considering what I was reasoning before.. Other times. Anne will probably only decide to switch if a < c/4. Bob: “Hmm. actually... we must simply take a closer look at the issue at hand. Conclusion Paradoxes provide is with an opportunity to grow in our understanding. Sometimes.

The Two-Envelope Paradox: A Complete Analysis?. References [1] Steven J. Marc Kilgour. Perspectives 3 (1989). which was then replaced with the superior ZermeloFraenkel set theory. seeking out the resolution of a paradox is an exciting way to bring our limited understanding closer in line with the pervasive consistency that underlies mathematics and our physical universe. In either case. Key College Publishing: Emeryville. Burger and Michael. 1st 171–181. [2] Edward B. Chalmers. CA [3] David illuminating the problems of so-called naive set theory. [5] Barry Nalebuff. The Two Envelopes Paradox MAA Online. Brams and D. 27–34. (1999).maa.html> . The Heart of Mathematics. [4] Keith Devlin. of Econ. Puzzles: The Other Person’s Envelope is Always Greener Jrnl. 12 . though. URL = <http://consc. URL = <http://www. The Box Problem: To Switch or Not to Switch Mathematics Magazine 68 (1995).html> .