The Two Envelopes Problem

There are several distinct diagnoses in the literature, and part of what keeps the problem alive is that they are answers to different versions of it.

Diagnosis 1: the equivocation. In the closed-envelope version, $A$ is a random variable that takes different values in the two branches, so $\mathbb{E}[\text{other}] = \tfrac12(2A) + \tfrac12(A/2)$ is not a valid computation of a conditional expectation. This is the standard textbook answer, and for the closed-envelope problem it is entirely correct. Nickerson and Falk call this the "same-symbol" error and show experimentally that readers who can identify it in the abstract still fall for it in the story.

Diagnosis 2: the impossible prior. The step "the other envelope is equally likely to be double or half, whatever I hold" requires a distribution over the smaller amount satisfying $\pi(a) = \pi(a/2)$ for all $a$ — flat over a geometric ladder, hence unnormalisable. It is an improper prior: a formal object that behaves like a distribution in some calculations and is not one. Improper priors are used deliberately and usefully in Bayesian statistics, and this problem is one of the cleanest demonstrations of what can go wrong when you do.

Diagnosis 3: the infinite expectation. Broome's 1995 construction is the uncomfortable one. Let the smaller amount be $2^{k}$ with probability $\tfrac13(\tfrac23)^{k}$ for $k = 0, 1, 2, \ldots$ This is a genuine, normalisable distribution — no impropriety anywhere. And under it, whatever amount you observe, the expected gain from swapping is strictly positive. The catch is that the distribution has infinite mean. David Chalmers argued in 2002 that this makes the surviving paradox a St Petersburg problem: the "always swap" advice is consistent but vacuous, because there is no finite expectation for it to improve on.

Diagnosis 4: it is not a paradox about probability at all. Some philosophers, John Norton among them, argue that the puzzle is best read as a demonstration of what happens when you apply an inference rule outside its domain of validity — closer to a division-by-zero error than to a genuine tension in probability theory. On this view "the paradox" is a family of distinct mistakes that happen to produce the same sentence.

What everyone agrees on. No proper prior makes swapping favourable in the closed-envelope case. Symmetry settles it: your envelope and the other are exchangeable random variables, so they have the same expectation whenever an expectation exists. Any argument concluding otherwise contains an error, and the only interesting question is which one.

The randomised-switching result. A genuinely surprising fact, due to Thomas Cover, and one that is not a fallacy. Suppose you open your envelope, see $a$, and use this rule: draw a threshold $T$ from any probability distribution with support on the whole positive line, and swap if $a < T$. Then your probability of ending up with the larger envelope is strictly greater than 1/2, for any pair of amounts the filler might have chosen. You cannot beat the envelopes on expected money, but you can beat them on the probability of getting the better one.

Why "debated". The mathematics of the closed-envelope case is settled and has been for decades. What remains contested is which diagnosis is the right one, whether the open-envelope Broome–Chalmers version constitutes a live paradox or a solved curiosity, and what the whole thing tells us about the legitimacy of improper priors in inference. Papers are still published on it.

1930s–1953. Maurice Kraitchik published "The Paradox of the Neckties" in his recreational-mathematics writing — first in French as Les Mathématiques des Jeux (1930), later in the English Mathematical Recreations, whose second edition appeared from Dover in 1953. He also gave a wallet variant, in which two men compare the contents of their wallets and the one with less takes the other's. Kraitchik states the contradiction plainly and offers no resolution.

1953. J. E. Littlewood published A Mathematician's Miscellany, which contains a card version of the same argument. Littlewood credits it to Erwin Schrödinger — the physicist, who had spent the war years in Dublin and was a compulsive puzzler. This is the earliest dated record tying the argument to a named originator, and it is why 1953 is the conventional date for the problem even though Kraitchik's version is older in substance.

1982. Martin Gardner presented Kraitchik's wallet version in Aha! Gotcha, naming it "the wallet game", and brought it to a general readership. Gardner did not resolve it either.

1988–1989. Barry Nalebuff posed the problem in the Puzzles column of the Journal of Economic Perspectives and, after what he described as an overwhelming volume of responses, devoted his 1989 column to it: "Puzzles: The Other Person's Envelope Is Always Greener". Nalebuff sorted the proposed solutions, drew the crucial distinction between the closed- and open-envelope versions, and gave the problem its envelopes. Everything written since uses his framing. A closely related statement appeared in 1988 in the Lindley and Smith volume.

1995. John Broome published "The Two-envelope Paradox" in Analysis, constructing the proper prior under which swapping is favourable at every observed value. This ended any hope that the problem could be dismissed as a simple algebraic slip.

2002. David Chalmers published "The St. Petersburg Two-Envelope Paradox", three pages arguing that the versions surviving the standard diagnosis are St Petersburg problems, and that the two paradoxes share a root in undefined expectations.

2006. Raymond Nickerson and Ruma Falk published a comprehensive review in Thinking & Reasoning, cataloguing the published resolutions, noting that several of them answer different questions, and analysing the cognitive pull of the flawed argument.

The literature has continued steadily since, in Analysis, Synthese, Theory and Decision and the statistics journals. The distribution of opinion has moved, but the problem has not closed.

The two envelopes problem is the best available demonstration that a probability calculation can be locally valid at every step and globally meaningless.

Each line of the 25% argument is the sort of move that appears in correct derivations every day: name a quantity, condition on two cases, weight by probabilities, add. Nothing looks wrong. The error is structural — a variable that quietly changes referent, and a probability assignment that no coherent belief state could hold.

Three consequences follow, and all three are practical.

Define your random variables before you condition on them. The equivocation in this problem is not exotic; it is what happens whenever a symbol is introduced mid-argument and then used as though it had been fixed in advance. This is the same failure that produces the base-rate fallacy, the prosecutor's fallacy and a large fraction of the errors in applied statistics.

Improper priors are tools, not free parameters. Flat priors over unbounded spaces are used constantly and usually harmlessly. This problem shows what happens when the impropriety is not recognised: the posterior recommendations become incoherent while every individual computation continues to look fine. Anyone using non-informative priors needs a version of this example in mind.

When the expectation is infinite, expected-value reasoning stops being advice. The Broome–Chalmers versions are not resolved by finding an error, because there is no error. They are resolved by recognising that "better in expectation" is not a meaningful recommendation when no expectation exists — the same lesson the St Petersburg paradox taught in 1738, arrived at from a completely different direction.

Bayesian statistics and default priors. Objective and reference priors are often improper, and the standard caution — check that the posterior is proper, and check that decisions derived from it are coherent — is taught with this problem as the cautionary example. It is the cleanest case of an improper prior producing confident nonsense.

Trading and market microstructure. "The other side of this trade looks better than mine" is the envelope argument, and it is wrong for the same reason: if both counterparties can run it, neither can be right. The winner's curse in auctions is the same structure with the equivocation made explicit — the fact that you won is information about the value, just as the fact that you are holding an envelope is information about which one it is.

Reinforcement learning and policy evaluation. Agents that re-evaluate an action using a value estimate conditioned on having already taken it can oscillate indefinitely between two options, each of which looks better from the other's vantage point. Preventing this is why off-policy evaluation is careful about which distribution an expectation is taken over.

Mechanism design and exchange protocols. Any protocol in which both parties must voluntarily agree to a swap has to survive the observation that a swap both parties consider favourable is usually impossible. The no-trade theorems of Milgrom and Stokey formalise exactly this, and the envelope paradox is their elementary illustration.

Teaching probability. The problem is now standard in graduate courses precisely because the fallacy is invisible to people who have already learned conditional expectation. It is a test of whether a student can say which distribution an expectation is taken over — which is the one thing the notation does not force you to state.

The flawed calculation, stated formally

Let $X$ be the amount in your envelope and $Y$ the amount in the other. The argument runs:

$$\mathbb{E}[Y] = \tfrac12(2X) + \tfrac12!\left(\tfrac{X}{2}\right) = \tfrac54 X$$

The error is that $\tfrac12$ is being used as $P(Y = 2X)$ unconditionally, while $X$ is being treated as a known constant. If $X$ is a constant, then $P(Y = 2X \mid X = x)$ must be evaluated at that $x$, and it is not $1/2$ in general. If $X$ is a random variable, then the expression is not a valid application of the law of total expectation, because the events ${Y = 2X}$ and ${Y = X/2}$ partition the space in a way that is correlated with the value of $X$.

Written with the correct partition, where $L$ is the event that you hold the smaller envelope:

$$\mathbb{E}[Y] = P(L),\mathbb{E}[2X \mid L] + P(L^{c}),\mathbb{E}!\left[\tfrac{X}{2} ,\middle|, L^{c}\right]$$

With $\mathbb{E}[X \mid L] = x$ and $\mathbb{E}[X \mid L^{c}] = 2x$, this gives $\tfrac12(2x) + \tfrac12(x) = \tfrac32 x = \mathbb{E}[X]$. The conditioning was the whole issue.

Why no proper prior makes swapping always favourable

Let the smaller amount be $X_{s}$ with prior density or mass $\pi$. You open your envelope and observe $a$. The pair is either $(a/2, a)$ or $(a, 2a)$, and since you picked an envelope at random,

$$P(\text{other} = 2a \mid \text{observe } a) = \frac{\pi(a)}{\pi(a) + \pi(a/2)}$$

For this to equal $1/2$ for every $a$ requires

$$\pi(a) = \pi(a/2) \quad \text{for all } a > 0$$

so $\pi$ is constant along the geometric chain ${a,2^{k} : k \in \mathbb{Z}}$. Any such $\pi$ has infinite total mass and cannot be normalised. It is an improper prior.

Improper priors are not always harmful — flat priors over the real line are standard in Bayesian analysis and often yield proper posteriors. But they carry no guarantee of coherence, and this problem is the textbook demonstration: reasoning that would be sound under a proper prior produces a Dutch-book-like recommendation under an improper one.

The finite-expectation theorem

Claim. If $\mathbb{E}[X_{s}] < \infty$, then swapping cannot be favourable for every observed value.

Proof sketch. Let $g(a)$ be the expected gain from swapping on observing $a$. Then $\mathbb{E}[g(A)] = \mathbb{E}[Y] - \mathbb{E}[X] = 0$ by exchangeability, provided both expectations are finite. A function with mean zero cannot be strictly positive everywhere on a set of full measure. So the set where swapping loses must have positive probability. $\square$

This is the precise sense in which "always swap" is impossible: it requires $\mathbb{E}[X] = \infty$, at which point the argument no longer establishes anything, because $\mathbb{E}[Y] - \mathbb{E}[X]$ is $\infty - \infty$.

Broome's distribution, worked through

Let the smaller amount be $2^{k}$ with probability

$$\pi_{k} = \tfrac13\left(\tfrac23\right)^{k}, \qquad k = 0, 1, 2, \ldots$$

These sum to $\tfrac13 \cdot \frac{1}{1 - 2/3} = 1$, so the prior is proper.

Observe $a = 2^{n}$ with $n \geq 1$. The pair is $(2^{n-1}, 2^{n})$ with weight $\pi_{n-1}$ or $(2^{n}, 2^{n+1})$ with weight $\pi_{n}$, each further weighted by the $1/2$ chance you opened the relevant envelope. The expected gain from swapping is proportional to

$$\pi_{n}\left(2^{n+1} - 2^{n}\right) + \pi_{n-1}\left(2^{n-1} - 2^{n}\right) = \pi_{n}2^{n} - \pi_{n-1}2^{n-1}$$

Substituting $\pi_{k}2^{k} = \tfrac13(\tfrac43)^{k}$:

$$\tfrac13\left(\tfrac43\right)^{n} - \tfrac13\left(\tfrac43\right)^{n-1} = \tfrac13\left(\tfrac43\right)^{n-1}\left(\tfrac43 - 1\right) = \tfrac19\left(\tfrac43\right)^{n-1} > 0$$

Strictly positive for every $n \geq 1$, and for $n = 0$ swapping is certain to gain. So swapping is favourable at every observed amount.

And the mean:

$$\mathbb{E}[X_{s}] = \sum_{k=0}^{\infty} \pi_{k}2^{k} = \sum_{k=0}^{\infty}\tfrac13\left(\tfrac43\right)^{k} = \infty$$

The ratio $4/3 > 1$ is doing all the work. The prior's tail decays like $(2/3)^{k}$ while the payoffs grow like $2^{k}$, so the expectation diverges. This is exactly the St Petersburg structure, and Chalmers' point is that the two paradoxes are the same paradox once you get this far.

Cover's randomised switching strategy

Suppose the filler chose amounts $x$ and $2x$ by any means whatever — you need no prior at all. Choose any random threshold $T$ with $P(T \in (u,v)) > 0$ for every $0 < u < v$; an exponential or lognormal draw will do. Open your envelope, see $a$, and swap if and only if $a < T$.

Let $p = P(x < T \leq 2x) > 0$ by construction. Then:

• If $T \leq x$: you never swap, so you end with the larger envelope with probability $1/2$.
• If $T > 2x$: you always swap, so again probability $1/2$.
• If $x < T \leq 2x$: you swap when holding $x$ and keep when holding $2x$ — you end with the larger envelope with probability 1.

Overall probability of ending with the larger envelope:

$$\tfrac12(1 - p) + 1 \cdot p = \tfrac12 + \tfrac{p}{2} > \tfrac12$$

The strategy works against any filler, adversarial or not, with no assumptions. It does not contradict the earlier results: it improves the probability of getting the larger amount, not the expected amount, and those come apart precisely because the two envelopes have different values.

Variants worth knowing

• The closed-envelope version. Fully resolved. The equivocation is the whole story, and symmetry forbids any advantage.
• The open-envelope version with a proper, finite-mean prior. Fully resolved. Swapping is favourable for some observed values and unfavourable for others, and averages to zero.
• The open-envelope version with an infinite-mean prior. This is the residue. "Always swap" is genuinely correct pointwise and genuinely useless globally, and what to make of that is where the philosophical disagreement lives.
• The symmetric two-player version (Kraitchik's neckties). Both players compute a positive expectation. Here the resolution is that the players' expectations are conditional on their own holdings and are not simultaneously satisfiable — the sum of the two gains must be zero.