“Paradox” is a useful invitation, but an imprecise diagnosis. An apparent contradiction may disappear once an ambiguous word is defined. A counterintuitive theorem may be completely consistent. A genuine logical antinomy may instead show that a collection of assumptions cannot all be retained. Keeping these categories separate helps us appreciate surprising mathematics without suggesting that ordinary arithmetic has stopped working.
Monty Hall: the host's procedure is the puzzle
There are three doors, one hiding a prize and two hiding goats. You choose one door. A host who knows all the contents always opens an unchosen goat door and always offers you the option to switch to the remaining closed door. Assume the prize was equally likely to be behind each door. Should you switch?
Yes. Your first choice wins with probability one third. It loses with probability two thirds. Under the stated procedure, switching wins exactly when your initial choice was wrong. The host's action does not turn your original choice into a fresh fifty-fifty selection. Instead, the informed removal of a goat channels the initial two-thirds chance into the one remaining alternative.
The assumptions are doing real work. If an uninformed host opens a random unchosen door and happens to reveal a goat, the conditional calculation changes. If a host offers a switch only in selected situations, their policy matters too. The familiar two-thirds answer belongs to a specified experiment, not to every story containing three doors. The problem is associated with the television host Monty Hall and became widely discussed after Marilyn vos Savant addressed it in print.
A coin puzzle about updated belief
A bag contains one fair coin and one coin with heads on both sides. You choose a coin uniformly at random and toss it once. The result is heads. What is the probability that you chose the double-headed coin? Assume each physical side of the fair coin is equally likely to land upward.
Two coins does not mean fifty-fifty
Hint: compare the likelihoods
The observed heads is guaranteed with one coin but happens only half the time with the other.
Reveal the calculation
The joint probability of choosing the double-headed coin and seeing heads is 1/2. The total probability of heads is (1/2 × 1) + (1/2 × 1/2) = 3/4. Divide 1/2 by 3/4 to obtain 2/3.
This is a compact example of Bayesian conditioning, named for Thomas Bayes. New evidence changes the relative plausibility of the two possible sources. The answer is not a statement that the selected coin has physically changed. It describes what can be inferred after observing a result under an explicit sampling process.
Hilbert's hotel: full is not always finite-full
Imagine a hotel with one room for each positive integer, all occupied. A new guest arrives. The manager asks the guest in room n to move to room n + 1. Every existing guest has a new room, and room 1 becomes available. Nothing similar works for a finite hotel with every room occupied.
Now suppose countably infinitely many new guests arrive, numbered 1, 2, 3, and so on. Move each existing guest from room n to room 2n. The even-numbered rooms hold the old guests; the odd-numbered rooms hold the new ones. This is a one-to-one assignment, not a physical construction plan. It demonstrates how a countably infinite set can have the same cardinality as one of its proper subsets.
The thought experiment is associated with David Hilbert and rests on ideas developed in Georg Cantor's study of infinite sets. It does not show that every infinity has the same size. Cantor's diagonal argument establishes that the real numbers cannot be listed in the same way as the positive integers. A picture of an endless corridor is an aid to reasoning, not a substitute for specifying the correspondence.
Simpson's paradox: averages carry hidden weights
A trend can hold separately in two groups and reverse when the groups are combined. Consider two hypothetical treatments. In an easier group, A succeeds in 9 of 10 cases and B in 80 of 100. In a harder group, A succeeds in 20 of 100 cases and B in 1 of 10. A has the higher success rate in each group: 90% versus 80%, and 20% versus 10%.
Combined, A succeeds in 29 of 110 cases, while B succeeds in 81 of 110. The reversal is possible because A was used mostly in the harder group and B mostly in the easier group. The pooled averages use different weights. These invented numbers illustrate the arithmetic; they are not clinical evidence or a recommendation about treatment.
Associated with statistician Edward H. Simpson, the paradox warns against deciding which aggregate comparison matters without asking a causal question. More data does not automatically erase selection effects. The relevant grouping depends on how the data were generated and what you are trying to estimate.
A better response to a paradox
When a result feels impossible, list the assumptions before rejecting the conclusion. What was chosen randomly? What information did an observer possess? Are you comparing equally weighted groups? Does “infinite” mean countably infinite? Once those questions are visible, the surprising result often becomes a precise lesson about the limits of a useful intuition.