๐Ÿงฎ Brain Teaser

The Lazy Prisoner: A Probability Paradox

Three prisoners, A, B, and C, each believe their chance of being pardoned is 13\frac{1}{3}. The warden knows that exactly one will be pardoned. Prisoner A asks the warden: "Please tell me the name of one of B or C who will NOT be pardoned." The warden agrees โ€” if both B and C are to be executed, he flips a fair coin to decide whom to name.

The warden says: "B will not be pardoned."

Question: What is the probability that A is pardoned, given this information? What about C?

Many people instinctively say A's probability jumps from 13\frac{1}{3} to 12\frac{1}{2} โ€” is this correct?

Bayes theoremconditional probabilityMonty Hallparadoxprisoner

Answer: The Lazy Prisoner: A Probability Paradox

Key Idea / Intuition

The warden's statement is not symmetric information about A and C. The warden was forced to name B if A is the one pardoned (since both B and C would be executed), but also sometimes names B when C is pardoned. Carefully tracking the warden's coin-flip behavior via Bayes' theorem reveals that A's probability stays at 13\frac{1}{3}, while C's probability jumps to 23\frac{2}{3}. The key insight: the information "B will not be pardoned" was almost guaranteed if C is the one pardoned (warden had no choice), but only had probability 12\frac{1}{2} of being said if A is pardoned. So hearing "B" is stronger evidence for C than for A.


Formal Proof / Solution

Setup. Let PP = event A is pardoned, QQ = event C is pardoned, RR = event B is pardoned. Each has prior probability 13\frac{1}{3}.

Let WBW_B = event that warden names B (says "B will not be pardoned").

Compute P(WBโˆฃeachย case)P(W_B \mid \text{each case}):

  • If A is pardoned: both B and C will be executed, so warden flips a fair coin.
    P(WBโˆฃP)=12P(W_B \mid P) = \frac{1}{2}

  • If B is pardoned: warden cannot name B (B is being pardoned), so must name C.
    P(WBโˆฃR)=0P(W_B \mid R) = 0

  • If C is pardoned: warden must name B (only non-pardoned candidate available, since A can't be named).
    P(WBโˆฃQ)=1P(W_B \mid Q) = 1

Apply Bayes' theorem:

P(WB)=P(WBโˆฃP)โ‹…13+P(WBโˆฃR)โ‹…13+P(WBโˆฃQ)โ‹…13P(W_B) = P(W_B \mid P)\cdot\frac{1}{3} + P(W_B \mid R)\cdot\frac{1}{3} + P(W_B \mid Q)\cdot\frac{1}{3}

=12โ‹…13+0โ‹…13+1โ‹…13=16+13=12= \frac{1}{2}\cdot\frac{1}{3} + 0\cdot\frac{1}{3} + 1\cdot\frac{1}{3} = \frac{1}{6} + \frac{1}{3} = \frac{1}{2}

Posterior for A:

P(PโˆฃWB)=P(WBโˆฃP)โ‹…P(P)P(WB)=12โ‹…1312=13P(P \mid W_B) = \frac{P(W_B \mid P)\cdot P(P)}{P(W_B)} = \frac{\frac{1}{2}\cdot\frac{1}{3}}{\frac{1}{2}} = \boxed{\frac{1}{3}}

Posterior for C:

P(QโˆฃWB)=P(WBโˆฃQ)โ‹…P(Q)P(WB)=1โ‹…1312=23P(Q \mid W_B) = \frac{P(W_B \mid Q)\cdot P(Q)}{P(W_B)} = \frac{1\cdot\frac{1}{3}}{\frac{1}{2}} = \boxed{\frac{2}{3}}

Conclusion:

| Prisoner | Prior | Posterior (after warden names B) | |----------|-------|----------------------------------| | A | 1/31/3 | 1/31/3 (unchanged!) | | B | 1/31/3 | 00 | | C | 1/31/3 | 2/32/3 |

Why does A not gain? A already knew that either B or C would be executed โ€” hearing which one adds no information about A's own fate. But it transfers all of B's probability mass to C, since C's execution was guaranteed to produce the answer "B." This is a classic instance of the Monty Hall phenomenon: the agent who had no choice is the one whose posterior surges.

The intuition that "now it's 50-50 between A and C" is the famous fallacy: it ignores the asymmetry in how the warden was constrained.

Source: Fifty Challenging Problems in Probability with Solutions, Frederick Mosteller (Problem 13: The Prisoner's Dilemma)

Type: ProbabilitySource: Fifty Challenging Problems in Probability with Solutions, Frederick Mosteller (Problem 13: The Prisoner's Dilemma)Edit on GitHub โ†—