Bertrand's Box Paradox

What Bertrand was actually arguing. The box paradox is not presented in Calcul des probabilités as a curiosity. It opens the book, and it opens it as a methodological warning: a probability problem is not well posed until you have said how the evidence was generated. The same volume contains the far more famous "Bertrand paradox" about the length of a random chord in a circle, where three different and perfectly reasonable readings of "at random" give three different answers — 1/2, 1/3 and 1/4. The two problems make the same point from opposite directions. In the chord case the procedure is genuinely ambiguous and no answer is uniquely correct. In the box case the procedure is specified, and readers still get it wrong.

The procedure matters here too. The 2/3 answer depends on a detail that is easy to skim past: you opened a drawer at random, and reported what was in it. Change that and the answer changes.

• Suppose instead someone who can see inside deliberately opens a drawer containing gold whenever one is available. Then seeing gold tells you only that your box is not the silver-silver box — the gold-gold and mixed boxes were both certain to produce this display — and the answer drops to 1/2.
• Suppose you are told merely "this box contains at least one gold coin", with no account of how anyone came to say that. Then the answer is genuinely indeterminate, and you have walked into the boy or girl paradox.

A dissenting note. Some commentators have argued that although Bertrand reached the right answer, his reductio is right for the wrong reason — that it establishes the inconsistency of the 1/2 answer without isolating the actual error, which is the failure to weight the boxes by how many gold coins they contain. The criticism is fair as far as it goes, but Bertrand's argument has an advantage the modern one lacks: it needs no formal machinery at all.

Why this counts as veridical. Nothing here is contradictory, and nothing dissolves under analysis. The reasoning is elementary, a formal demonstration exists, and the surprising answer is simply the true one. That is Quine's definition of a veridical paradox, and Bertrand's box is about as clean an example as exists.

1889. Joseph Bertrand published Calcul des probabilités with Gauthier-Villars in Paris. He was 67, permanent secretary of the Académie des Sciences, and had spent decades teaching at the École Polytechnique and the Collège de France. The box problem appears near the very start, in his original framing as three identical chests each with two drawers, each drawer holding one coin.

Bertrand posed it, gave the 2/3 answer, supplied the reductio, and added the 300-box illustration. There is no mystery about the origin of this one: it is signed, dated and solved in a single place.

1959. Martin Gardner published the Three Prisoners problem in Scientific American. Three prisoners, one to be pardoned; A asks the warden to name one of the other two who will be executed, and the warden names B. Does A's chance improve from 1/3 to 1/2? It does not — it stays at 1/3, while C's rises to 2/3. This is Bertrand's box with the coins replaced by pardons and a warden supplying the observation. Gardner's readers reacted much as Bertrand's had.

1975. Steve Selvin posed the same structure again in terms of a television game show, and it acquired the name it now carries. The Monty Hall problem has three objects, one of which is revealed by a rule-following agent, and probability 2/3 landing on the option you did not initially choose. Bertrand got there 86 years earlier.

The line of descent is real, not decorative. Boxes with two drawers, then prisoners and a warden, then doors and a host: the same conditional structure, restated three times, each time provoking the same public disbelief. That recurrence is the most interesting historical fact about the problem. Each generation encountered it as new.

Bertrand's box is the ancestor of the most argued-about puzzle in probability, and it is much the better teaching example — because it strips away everything that makes Monty Hall contentious.

With Monty Hall, arguments stall on the host's motives. Does he always open a door? Does he know where the car is? Would he have offered the switch if you had picked differently? Those questions are legitimate and they are also a distraction, because they let a reader who has not grasped the conditioning believe the disagreement is about game-show trivia.

Bertrand's boxes have no host. There is nobody to have motives. The drawer opens at random. And the answer is still 2/3, still for the same reason. Whatever is going on, it is not about Monty.

The transferable lesson is a rule about evidence:

When evidence could have arisen in several ways, weight each hypothesis by how many ways it could have produced exactly what you saw — not by how many hypotheses survive.

Counting surviving hypotheses gives 1/2. Counting the ways they could have generated your observation gives 2/3. That difference is the entire content of the likelihood term in Bayes' theorem, and Bertrand demonstrated it decades before anyone was routinely teaching Bayes' theorem to undergraduates.

Teaching Bayesian reasoning. The box problem is standard in introductory probability courses precisely because the sample space is small enough to write out completely. A student who can enumerate six equally likely rows and cross out three of them has performed a Bayesian update without needing the formula, and can then be shown that the formula agrees.

Diagnostic testing. A positive result should be weighted by how likely each condition was to produce that particular result — not merely by which conditions remain possible. Clinicians shown "the test is 99% accurate" routinely give the box-paradox answer, and the corrective is the same: count the ways the observation could have arisen.

Forensic and genetic evidence. A DNA profile, a fibre match or a partial fingerprint is evidence whose strength depends on how many sources could have generated it. The prosecutor's fallacy — treating "this evidence is unlikely if innocent" as "innocence is unlikely given this evidence" — is the 1/2 answer in a courtroom.

Machine learning and class imbalance. A classifier's posterior depends on both the likelihood and the prior. Models trained on rebalanced data implicitly change the prior, and calibrating them back is the same arithmetic as re-weighting Bertrand's boxes by their coin counts.

Anywhere a system reports partial state. A monitoring system that says "at least one replica is unhealthy" tells you something different depending on whether it checked a random replica or searched for an unhealthy one. That is Bertrand's distinction between a random drawer and a deliberately chosen one, and it is why observability tooling documents its sampling policy.