The Other Envelope

  • Hard
  • Logical puzzles
  • Tests: Finding the flaw in an expected-value calculation

The puzzle

Two envelopes contain money, one holding exactly twice the other. You pick one and find £100. You may keep it or swap.

The argument for swapping: the other holds £50 or £200 with equal chance, so its expected value is £125 - better than £100. But the same argument applies to whichever envelope you hold, so you should swap for ever. Where is the error?

Stuck?

One at a time, smallest first. Open only as many as you need — the whole point of the puzzle is the part you work out yourself.

Hint 1

The calculation treats "£50 or £200, equally likely" as given. Is it?

Hint 2

Ask what the two envelopes actually contain. There are only two possible pairs, and they cannot both be in play.

Show the answer Hide the answer

The error is assuming the other envelope is equally likely to be half or double. Once the pair is fixed, seeing £100 is evidence about which pair you are in, and the two cases are not generally equally likely.

How to get there

The envelopes hold either (£50, £100) or (£100, £200). These are different situations, not two outcomes within one situation. The 50/50 split is assumed, not derived.

For it to hold for every amount you might see, the amounts would have to be drawn from a distribution that is equally likely to produce any pair, however large - and no such distribution over the positive numbers exists. The calculation quietly relies on an impossible prior.

Set it up honestly and the paradox disappears. If the smaller amount is x, the envelopes hold x and 2x, your envelope averages 1.5x, and so does the other: swapping gains nothing. The £125 came from using "£100" to mean the smaller amount in one branch and the larger in the other, which is two different quantities wearing one label.

The technique worth keeping

When an expected-value argument proves something absurd, check whether the same symbol is standing for two different quantities across the branches. That, and an assumed-uniform prior over an unbounded range, account for most paradoxes of this shape.