The Number You Made Up

Here is the argument that will not leave you alone. You hold an envelope with some amount X inside. The other envelope holds either 2X or X/2 — a coin-flip, you tell yourself, so its expected value is ½·2X + ½·(X/2) = 1.25X. That's a quarter more than what you hold. Switch. And since the same sum works no matter what X turned out to be, you should switch before you even open it — and then switch back, forever, envelope to envelope, richer each time. Something is plainly, deeply broken. But where?

Run it long enough and the wandering average settles onto 0. Not a small number — zero, exactly, and for any distribution the amounts could be drawn from. The reason is almost insultingly simple: the envelope handed to you is a uniformly random one of the two, so on average what you hold and what you'd switch to are the same thing. Switching is a coin you paint gold and back again. There is no free lunch on this table.

So the 1.25X calculation is lying — but it looks airtight. Its single false word is coin-flip. The step "the other envelope is 2X or X/2 with equal odds, whatever X I see" cannot be true for every X at once. To believe it for all values is to believe in a distribution that spreads equal weight across every scale from a penny to a googol of pounds — a thing that cannot be normalised, an infinite probability that sums past one. Assume a real, honest prior instead and the equal-odds premise breaks the moment you look. Watch it break.

There it is, laid flat. At every value in the interior the bar is green: switching really does gain, on average, a quarter of what you're looking at — the "+25%" was never a hallucination. And at the very bottom value switching is a guaranteed win. But look at the single red bar on the right: the largest amount you could open, the one where the other envelope must be the small one and switching halves a small fortune. That one bar, weighted by how rarely you see it, is exactly as heavy as all the green ones combined. The paradox isn't that the gains are fake. It's that the argument quietly drops the one catastrophic value off the edge of the table — the value that only exists because a real distribution has to stop somewhere. Let it stop, and the books balance to zero, in exact arithmetic, every time.

The naive sum is right about every value except the last one it pretends doesn't exist — and the last one weighs as much as all the rest.

Now the trick that's real

So you cannot beat this game by thinking harder about your own envelope. Fine. Here is a different, stranger question, and it has a genuinely astonishing answer. Forget doubling; forget priors entirely. An adversary — someone who wants you to lose — writes any two different numbers on two slips, seals them, and hands you one at random. You look. Can you decide whether the number in your hand is the larger of the two, and be right more than half the time?

Every instinct says no. You've seen one number with no scale, no context, nothing to compare it to. Half the time it's the big one, half the time the small one, and a single naked number can't tell you which — you should be stuck at a coin flip forever. Except you're not. In 1987 the information theorist Thomas Cover gave a method that provably beats the coin, against any adversary, with no knowledge of the numbers at all. The method is this: before you look, invent a number. Draw a threshold T from any distribution you like — pull it out of the air — as long as it has some chance of landing anywhere on the line. Then compare. If your envelope beats your invented number, keep it. If it doesn't, switch.

Trace what just happened. Your invented number T lands in one of three places. If it falls below both hidden amounts, your envelope always beats it, so you always keep — and you're right exactly half the time, the coin you started with. If it falls above both, you always switch, and again you're right half the time. But if T happens to land between the two amounts — and it has some chance to, because you gave it a chance to land anywhere — then the comparison stops being noise and becomes an oracle. Your envelope beats T if and only if it's the larger one. In that slice of luck you are right with certainty. The middle region is the entire edge, and it is worth exactly half its own probability:

P(you pick the larger) = ½ + p₂ / 2,  where p₂ is the chance your invented number lands between the two you were never told.

A number you pulled from nowhere, that knows nothing about the game, still carries information — because where it falls relative to your envelope is real, checkable news, and a positive fraction of the time that news is decisive. This is the whole shy miracle of Cover's result: randomness you author yourself can be a source of genuine knowledge.

Two honest fences around it. First — the edge is real but the adversary owns its size. Press hides in the tail above: they can shove both amounts far out where your threshold almost never reaches, and your advantage, while it stays strictly above ½, shrinks toward a whisper. There's no universal floor on how large the edge is without some hint about scale; there is only the guarantee that it never drops to zero. Second — this is a promise about the probability of grabbing the larger envelope, not about expected money in the prior-free game. Those are different questions, and the honest work keeps them apart; the money story is the one Part I settled, and it needs a real prior to even ask.