Alan Smithee wrote: “I thought SSitter was right that december was being deliberately obscurant by ignoring what was clearly an unspoken assumption in the problem, that all factors not mentioned are deemed irrelevant.”
Cecil Adams discusses another problem and includes in his answer the words:
“Which box would you choose, given the information that you have?”
In the 2 envelope problem there are two relevant factors, the donor’s decision of how much money to put in the envelopes, and our randomised choice of an envelope. Following the suggestions above, we would like to answer the problem “given the information that you have.” That is, following the “unspoken assumption in the problem that all factors not mentioned are deemed irrelevant.”
The key to this problem, I believe, is that it is mathematically impossible for the donor’s choice to be irrelevant. Here’s why:
Let’s assume that the donor chose a method of putting cash in the envelopes so that his decision is irrelevant. Specifically, we assume that the donor assigned amounts of money to the envelopes so that after you chose an envelope and opened it, it would always be 50-50 as to whether the other envelope had half that amount or twice that amount, regardless of how much money you found in the envelope.
We will show that this assumption leads to a contradiction. There is no mechanism leading to this result! That’s why the apparent paradox (described in an earlier post) exists.
To put this assumption in Bayesian terms, we are looking for a prior distribution of the original envelopes such that after you choose an envelope at random and open it, it will be 50-50 as to whether the amount in the other envelope is larger or smaller.
I am asserting that no such prior distribution exists.
To see why, suppose the envelope you opened had $4. Your assumption says that it should be equally probable as to whether the other envelope has $2 or $8. That happens if you were equally likely to be given (2,4) or (4,8) (where the figures in parentheses indicate the amount of money in the two envelopes). In Bayesian terms, the prior probabilities of (2,4) and (4,8) must have been equal.
For example, you could assume that the donor prepared just two sets of envelopes, (2,4) and (4,8), and chose one of the pairs at random. Then, when you see $4 in the envelope you opened, it would indeed be 50-50 as to whether the other envelope has $2 or $8. So far so good.
However, remember we want our assumption to hold regardless of how much money you found in the envelope. This is where the problem begins.
Once you assume that the donor chose from (2,4) and (4,8), then he might have given you the (4,8) and you might have opened the $8 envelope. So, you must also assume that the prior distribution had equal probabilities for (8,16) and (4,8). E.g., you might assume that the donor chose randomly from 3 sets of envelopes, (2,4), (4,8) and (8,16).
However, by similar reasoning, there would need to be equal probabilities for (16,32), (32,64), etc. Even ignoring the limitation on amounts of money, this argument leads to the conclusion that in the prior distribution a countably infinite number of possibilities all would need equal probability, call it P.
But, if P > 0, then the total probability would be infinity, whereas it must be one.
And if P = 0, then the total probability equals 0, which is also a contradiction.
This contradiction shows that you cannot satisfy the assumption that after you opened your chosen envelope it would always be 50-50 as to whether the other envelope has half that amount or twice that amount,
regardless of how much money you found in the envelope.
I must admit that you could satisfy the above assumption substituting the word “sometimes” for “always.” You could assume, e.g., that the donor chose randomly from (2,4) and (4,8). When you opened the envelope with $4, it was 50-50 as to whether the other envelope had $2 or $8.
But, if you had opened an envelope with $2 or $8, then the odds would NOT have been 50-50.
IMHO, this assumption destroys any interest in the problem. I believe that the problem as stated wanted to include the assumption, above, but it cannot be fulfilled.
Finally, I must confess that my “proof” above is not written with perfect mathematical rigor, but I do believe that it is correct.
P.S. It was asked what L. Savage said.
He discussed how probability could be defined when the frequentist approach did not apply, say, because one was considering a unique, one-time event rather than repeated trials. One key point of his brilliant book is that a judgmental probability can be consistently defined
as personal opinion. The definition of this probability could be deduced in theory from which side a person would take in a series of bets. A person’s subjective probability would obey the normal mathematical laws of probability.
However, there would be no contradiction if two different people assigned different
probabilities to the same event.