Fact-checked by Grok 2 weeks ago

Three prisoners problem

The Three prisoners problem is a classic probability paradox that demonstrates the counterintuitive nature of and information revelation. First introduced by in his 1959 Scientific American column on mathematical games, it involves three s (A, B, and C) sentenced to death, with one secretly pardoned at random by the governor, and a warden who, upon request from prisoner A, names another prisoner doomed to die, prompting a reevaluation of survival odds that defies initial intuition. In the setup, the pardon is equally likely for any of the three prisoners, giving each an initial 1/3 probability of survival. Prisoner A, unaware of the outcome, asks the warden—who knows the pardoned identity—to reveal the name of one other prisoner (B or C) who will be executed, with the stipulation that if both B and C are doomed (i.e., if A is pardoned), the warden selects randomly between them with equal probability. The warden responds that B will die. At this point, the problem poses: what is the updated probability that A is the pardoned prisoner? The paradox emerges from the intuitive but incorrect reasoning that the revelation eliminates B, leaving A and C as equally likely candidates for pardon, thus raising A's survival chance to 1/2. In reality, the that A is pardoned remains 1/3, while C's jumps to 2/3; this holds because the warden's choice is not independent of the underlying pardon but always reveals a doomed other than A, skewing the probabilities without providing new information about A's fate specifically. To resolve it formally, consider the of equally likely pardon outcomes: if C is pardoned (probability 1/3), the warden must name B; if B is pardoned (1/3), the warden must name C; if A is pardoned (1/3), the warden names B or C each with probability 1/2. on the warden naming B yields four scenarios with probabilities 1/3 (C pardoned), 1/3 (B pardoned, but impossible here), 1/6 (A pardoned, names B), and 0 (A pardoned, names C, but not this case)—normalizing gives A's probability as (1/6) / (1/3 + 1/6) = 1/3, and C's as 2/3. The problem is structurally analogous to the later-popularized , serving as an earlier variant that highlights errors in probabilistic reasoning under partial information revelation, and it has been widely analyzed in probability education to teach Bayesian updating and the importance of defining events carefully.

Background and Problem Statement

Historical Origins

The three prisoners problem originated as a posed by in the October 1959 issue of his "Mathematical Games" column in , where it was presented as a counterintuitive probability puzzle involving conditional . , a prolific and skeptic who authored over 70 books, played a pivotal role in popularizing through his 25-year tenure at from 1956 to 1981, introducing puzzles that bridged amateur enthusiasts and professional mathematicians while emphasizing and intellectual play. His column reached a wide , fostering interest in mathematical curiosities and influencing generations of readers to engage with abstract concepts in an accessible manner. The puzzle received its first formal analysis and explicit naming as the "three prisoners problem" within probability literature in the early 1960s, notably appearing in Frederick Mosteller's 1965 collection Fifty Challenging Problems in Probability with Solutions, where it was adapted to illustrate Bayesian updating and the pitfalls of intuitive probability assessment. This academic treatment solidified its status as a example in statistical education. Over subsequent decades, the problem evolved in its presentation across probability textbooks and pedagogical resources, often reframed to emphasize and information revelation, thereby influencing the teaching of these topics by providing a memorable that highlights common misconceptions in probabilistic reasoning.

Detailed Setup and Query

In the three prisoners problem, three prisoners labeled A, B, and C are held in separate cells under sentence of death. The governor selects one of them at random to pardon, with each having an equal probability of 1/3, by placing their names on slips of paper in a and drawing one secretly; the other two are to be executed. The warden, who knows the identity of the pardoned prisoner, is instructed to keep this information confidential for several days. Prisoner A, having heard a rumor of the impending , asks the warden to name one other prisoner who will be executed. The warden agrees under the following protocol: if B is pardoned, the warden must name C; if C is pardoned, the warden must name B; and if A is pardoned, the warden will randomly choose to name either B or C by flipping a . This ensures the warden always truthfully identifies at least one non-pardoned prisoner other than A, without revealing whether A is pardoned. The next day, the warden informs A that B will be executed. At this point, A wonders whether this revelation changes his own chances of being pardoned, initially believed to be 1/3, or if it now implies a 1/2 probability between A and C surviving.

Core Solution

Probabilistic Formulation

The probabilistic formulation of the three prisoners problem employs to model the uncertainty surrounding the pardon decision and the guard's revelation. Define the events P_A, P_B, and P_C as the events that prisoners A, B, and C are pardoned, respectively. The prior probabilities are uniform, with P(P_A) = P(P_B) = P(P_C) = \frac{1}{3}, reflecting the random selection of one prisoner for pardon. Let G_B denote the event that the guard names prisoner B as one to be executed when queried by prisoner A. The guard's behavior is formalized under the following assumptions: the guard always names a prisoner other than A who is not pardoned; if both B and C are to be executed (i.e., under P_A), the guard selects B with probability \frac{1}{2}. Thus, the conditional probabilities are P(G_B \mid P_A) = \frac{1}{2}, P(G_B \mid P_B) = 0 (since B cannot be named if pardoned), and P(G_B \mid P_C) = 1 (since B must be named if C is pardoned). The objective is to compute the posterior probability P(P_A \mid G_B), which represents prisoner A's updated probability of being pardoned given that the guard names B. This is obtained via : P(P_A \mid G_B) = \frac{P(G_B \mid P_A) P(P_A)}{P(G_B)}, where the marginal probability P(G_B) is given by the : P(G_B) = P(G_B \mid P_A) P(P_A) + P(G_B \mid P_B) P(P_B) + P(G_B \mid P_C) P(P_C). Analogous expressions apply for the posteriors P(P_B \mid G_B) and P(P_C \mid G_B).

Step-by-Step Calculation

To derive the posterior probabilities using , first compute the likelihood P(G_B \mid P_A), the probability that the names B given that A is pardoned. In this case, both B and C are to be executed, so the selects one at random to name, yielding P(G_B \mid P_A) = \frac{1}{2}. Next, compute P(G_B \mid P_B), the probability that the names B given that B is pardoned. Here, the cannot name B (the pardoned ) and also avoids naming A (the asker), so must name C instead, resulting in P(G_B \mid P_B) = 0. Then, compute P(G_B \mid P_C), the probability that the guard names B given that C is pardoned. With B to be executed and the guard avoiding naming A, B is the only option, so P(G_B \mid P_C) = 1. The prior probabilities are equal: P(P_A) = P(P_B) = P(P_C) = \frac{1}{3}. The marginal probability P(G_B) is found using the : P(G_B) = P(G_B \mid P_A) P(P_A) + P(G_B \mid P_B) P(P_B) + P(G_B \mid P_C) P(P_C) = \left( \frac{1}{2} \right) \left( \frac{1}{3} \right) + (0) \left( \frac{1}{3} \right) + (1) \left( \frac{1}{3} \right) = \frac{1}{6} + 0 + \frac{1}{3} = \frac{1}{2}. Applying gives the posterior P(P_A \mid G_B) = \frac{P(G_B \mid P_A) P(P_A)}{P(G_B)} = \frac{ \left( \frac{1}{2} \right) \left( \frac{1}{3} \right) }{ \frac{1}{2} } = \frac{1/6}{1/2} = \frac{1}{3}. Similarly, P(P_C \mid G_B) = \frac{P(G_B \mid P_C) P(P_C)}{P(G_B)} = \frac{ (1) \left( \frac{1}{3} \right) }{ \frac{1}{2} } = \frac{1/3}{1/2} = \frac{2}{3}, and note that P(P_B \mid G_B) = 0 since the guard named B to be executed. These posteriors sum to 1 (\frac{1}{3} + 0 + \frac{2}{3} = 1), confirming consistency.

Building Intuition

Scenario Enumeration

To build intuition for the Three prisoners problem, consider the possible underlying realities—or "worlds"—in which one is pardoned, each equally likely with \frac{1}{3}. These worlds are: (1) A is pardoned (B and C executed), (2) B is pardoned (A and C executed), or (3) C is pardoned (A and B executed). In each world, the warden, when asked by A to name another who will be executed, follows the rule of naming one who is not pardoned; if both B and C are to be executed (i.e., A pardoned), the warden chooses randomly between them with equal probability \frac{1}{2}. In the first world (A pardoned, probability \frac{1}{3}), the warden names B with probability \frac{1}{2} (subcase probability \frac{1}{3} \times \frac{1}{2} = \frac{1}{6}) or C with probability \frac{1}{2} (subcase probability \frac{1}{6}). In the second world (B pardoned, probability \frac{1}{3}), the warden must name C (since B cannot be named as executed), yielding probability \frac{1}{3} for this subcase. In the third world (C pardoned, probability \frac{1}{3}), the warden must name B, yielding probability \frac{1}{3} for this subcase. Suppose the warden names B as to be executed. The relevant scenarios are then those where the warden names B: the subcase of A pardoned and warden names B (probability \frac{1}{6}), and the case of C pardoned and warden names B (probability \frac{1}{3} = \frac{2}{6}). The total probability of these scenarios is \frac{1}{6} + \frac{2}{6} = \frac{3}{6} = \frac{1}{2}, so the that A is pardoned given the warden names B is \frac{\frac{1}{6}}{\frac{1}{2}} = \frac{1}{3}. This enumeration aligns with the core probabilistic result that A's chance remains \frac{1}{3}. The following table summarizes the scenarios, their probabilities, and the conditional outcomes when the warden names B:
ScenarioPardoned PrisonerWarden NamesProbability of ScenarioProbability Given Warden Names B
1aAB\frac{1}{6}\frac{1}{3} (for A pardoned)
1bAC\frac{1}{6}N/A (names C, not B)
2BC\frac{1}{3}N/A (names C, not B)
3CB\frac{1}{3}\frac{2}{3} (for C pardoned)
This discrete breakdown highlights how the warden's response filters the possible worlds without altering the relative likelihoods in the conditioned sample space.

Addressing Initial Intuitions

Initially, prisoners A, B, and C each believe they have an equal 1/3 probability of being pardoned, as the selection is random. When the guard informs A that B will be executed, the immediate intuition is that the pardon must now be equally likely between A and C, shifting their probabilities to 1/2 each due to apparent symmetry among the remaining candidates. This symmetry fallacy arises from overlooking the guard's knowledge and selective revelation, treating the information as simply eliminating B without adjusting for conditional dependencies. Enumeration of scenarios clarifies this by explicitly listing the possible pardon outcomes and the guard's responses, revealing that the case where B is pardoned is indeed eliminated, but the scenario where A is pardoned only contributes to half of the instances where the guard names B—due to the guard's random choice between B and C in that situation. In contrast, if C is pardoned, the guard always names B. As shown in the enumerated cases, this results in A retaining a 1/6 probability in the relevant branch versus 1/3 for C, making C's overall probability twice as likely after the revelation. This approach highlights how the initial equal probabilities are unevenly redistributed by the guard's action. The guard's naming of B provides asymmetric evidence: it would occur with certainty if C is pardoned but only with 50% probability if A is pardoned, favoring C as the more likely recipient of the pardon. This informational asymmetry underscores that the guard's statement is not neutral but conditioned on the true pardon, altering the posterior probabilities without changing the prior for A. A simple analogy is to an uneven branching , where the path leading to the guard naming B has a thicker branch from the C-pardoned root (full probability) compared to the thinner branch from the A-pardoned root (half probability due to ), illustrating why the probabilities do not split evenly.

Unpacking the Paradox

Sources of Misconception

One common misconception in the Three Prisoners problem arises from treating the guard's of prisoner B's fate as a simple elimination that equally divides the remaining probability between A and C, leading to an erroneous conclusion of 1/2 probability for A. This error ignores the conditional nature of the information provided by the guard, who knows the pardoned and deliberately chooses B based on a specific rather than randomly selecting among the non-pardoned. As a result, the revelation does not symmetrically reduce the possibilities but skews the probabilities due to the guard's constrained choice mechanism. Another frequent error involves assuming symmetry between prisoners A and C after the guard names B, overlooking how the guard's strategy makes naming B more probable if C is pardoned than if A is. Under this flawed view, individuals equate the two remaining prisoners as equally likely, failing to account for the fact that the guard would have named C instead if B were pardoned, thereby breaking the apparent balance. This assumption persists because it aligns with intuitive notions of fairness in random selection but neglects the non-random, knowledge-based decision of the guard. A third misconception stems from conflating the initial prior probabilities of 1/3 for each prisoner with the posterior probabilities after the guard's response, without properly updating via . People often reason that since two prisoners remain uncertain, the chances must now be 1/2 each, as in a example: "Only A and C are left, so it's 50-50 between them." This argument fails because it does not incorporate the new evidence from the guard's deliberate action, which alters the likelihoods asymmetrically—the correct posteriors being 1/3 for A and 2/3 for C. Psychological factors exacerbate these errors, particularly the tendency to neglect the role of the guard (analogous to in related puzzles) due to base-rate neglect and other cognitive es. Studies on in such probability puzzles show that individuals often underweight the informational value of the figure's constrained response, prioritizing initial equal priors over Bayesian updating. This reflects a broader difficulty in handling conditional probabilities, where the structured process is intuitively dismissed in favor of simplistic equiprobability.

Key Insights and Resolutions

The counterintuitive result that A's probability of pardon remains 1/3 after the guard names B stems from the dependent nature of the guard's action. Unlike an independent revelation that might equalize the remaining probabilities, the guard's selection rule—naming a non-pardoned other than A when possible—transfers the probability mass originally associated with B directly to C, without altering A's initial 1/3 share. This transfer occurs because, conditional on A being pardoned, the guard chooses randomly between B and C, but when B is pardoned, the guard must name C, skewing the conditional probabilities accordingly. A useful consistency check reinforces this: if prisoner A were to switch his request for information to prisoner C (effectively betting on C's ), the probability that C is pardoned would then be 2/3, mirroring the elevated chance assigned to C in the original setup and validating the asymmetric . This symmetry in the "switching" outcome aligns with the formal calculations, demonstrating that the model's predictions hold under alternative strategies. The paradox largely arises from underappreciating the nuances of , where intuitive symmetry assumptions overlook how the guard's partial information updates beliefs unevenly; resolving it highlights Bayes' theorem's role in rigorously incorporating the source's constrained behavior to compute posteriors, such as P(A pardoned | guard names B) = 1/3. A broader lesson is that solving such puzzles requires explicitly modeling the information provider's decision process, as failing to do so leads to incorrect equiprobability inferences about the remaining options. Empirical validation through simulations, running thousands of trials mimicking the pardon selection and guard's revelation, consistently yields approximately 1/3 success rate for prisoner A's pardon, affirming the theoretical resolution over repeated iterations.

Connections and Extensions

Similar Probability Puzzles

The , first formulated by statistician Steve Selvin, presents a scenario where a contestant selects one of three doors hiding either a car or , after which the host, who knows the contents, reveals a behind one of the unselected doors; switching to the remaining door yields a two-thirds probability of winning the car. This puzzle bears a direct analogy to the three prisoners problem, as the host's deliberate revelation of a non-prize door parallels the guard's informed naming of a non-pardoned prisoner, both updating conditional probabilities in a counterintuitive manner. The , introduced by , involves a family with two children where at least one is a boy; the probability that both are boys is one-third under standard assumptions of equal likelihood and independence. This shares structural similarities with the three prisoners problem through the enumeration of equally likely outcomes and the conditional update based on partial information, highlighting how specifying "at least one" alters the . Bertrand's box paradox, posed by , features three boxes—one with two gold coins, one with two silver coins, and one with one of each—where drawing a gold coin leads to a two-thirds posterior probability that it came from the mixed box. Like the three prisoners problem, it demonstrates Bayesian updating from an initial uniform upon observing that correlates with multiple hypotheses, revealing non-intuitive shifts in probabilities. The Sleeping Beauty problem, developed by Adam Elga, asks for the probability that a landed heads when Beauty awakens repeatedly under heads but only once under tails, with each time; arguments center on self-locating beliefs yielding one-third for heads. It ties loosely to the three prisoners problem by involving probabilistic updating amid uncertainty about one's position in a structured scenario, though it emphasizes epistemic rather than evidential revelation. These puzzles all feature non-intuitive conditional probability updates from seemingly symmetric setups, but the three prisoners problem uniquely stresses the role of an informed third party in providing targeted information that redistributes probabilities without altering the underlying priors.

Broader Applications and Variations

The three prisoners problem exemplifies updating in , particularly in scenarios involving partial information revelation. This structure parallels , where partial observations revise probabilities. In , the problem informs in Bayesian networks, enabling agents to revise probabilities upon receiving partial observations, as demonstrated in security applications where incomplete disclosures affect threat assessments. Pickering and Xu (2008) analyze the problem to expose pitfalls in network modeling, such as mishandling conditional dependencies, which can lead to erroneous belief updates in systems with hidden variables. Variations extend the core setup, including a four-prisoner version where one of four is pardoned at random, and the guard, upon query, names one of the remaining three to be executed (choosing randomly if multiple options exist), resulting in the querying prisoner's survival probability remaining 1/4 while the unnamed one's rises to 3/4. The problem generalizes to n prisoners with k pardons, where the guard announces m executions from the non-pardoned group; in this framework, the probability for the querying prisoner stays at k/n, but announcements can shift odds for others depending on selection rules. Wechsler et al. (2003) formalizes this extension, showing how neutrality in guard choices preserves initial probabilities for the querier while informativeness varies with n and k. Computational simulations, often via methods, empirically validate these probabilities by running thousands of trials of pardon assignments and guard responses, confirming the counterintuitive shifts (e.g., 1/3 vs. 2/3 in the classic case) and aiding visualization of conditional dependencies. In , this approach supports handling partial observations in models like partially observable Markov decision processes (POMDPs), where agents infer hidden states from revelations akin to the guard's statement, enhancing under in sensor management or . Extensions include quantum variants, where superposition allows strategies that outperform classical limits; for instance, the quantum Monty Hall problem—mathematically equivalent to the three prisoners setup—enables a player to achieve up to 100% success probability by entangling choices, shifting odds through measurement-induced collapses. D'Ariano et al. (2002) demonstrate how quantum operations on the "doors" (or prisoners) introduce non-local correlations, altering conditional probabilities beyond deterministic revelations. Adversarial guard strategies, where the guard deliberately selects revelations to minimize the querier's updated odds, further modify probabilities; in such cases, the querier's survival chance can drop below the neutral 1/3 if the guard biases toward high-impact disclosures, underscoring the role of information source intent in Bayesian updates. Educationally, the problem has been integrated into probability curricula since the 1970s to illustrate conditional reasoning, with empirical studies revealing persistent student misconceptions, such as assuming equiprobability among remaining options post-revelation (e.g., judging both the querier and named survivor as 1/2 likely). Rubel (2006) reports that middle and high school students (grades 5–11) exhibited this bias in about 57% of responses to an isomorphic three cards scenario, mirroring errors in related puzzles and linking to broader intuitive failures in partitioning sample spaces. Complementary research on the equivalent Monty Hall problem identifies additional fallacies, like representativeness heuristic overuse, informing targeted interventions to foster accurate probabilistic intuition. The analogy to the Monty Hall problem highlights its pedagogical value in dispelling such errors.

References

  1. [1]
    Mathematical Games: Probability and Ambiguity - Scientific American
    The answer to the problem of the three prisoners is that A's chances of being pardoned are 1⁄3 and that B's chances are 2⁄3. Regardless of who is pardoned ...
  2. [2]
    [PDF] The Monty Hall Problem Yurixy Lopez Martinez A project presented ...
    May 10, 2019 · It says that there are three prisoners, A, B and C, in separate cells and sentenced to death.
  3. [3]
    Monty Hall Problem - Penn Math
    An earlier version, the Three Prisoner Problem, was analyzed in 1959 by Martin Gardner in the journal Scientific American. He called it "a wonderfully ...Missing: source | Show results with:source
  4. [4]
    Martin Gardner, Genius Of Recreational Mathematics - NPR
    Apr 12, 2014 · So he really did bring out to the non-mathematician the clever mathematical ideas that mathematicians had already been able to observe in ...Missing: popularizing | Show results with:popularizing
  5. [5]
    Recreational Computing | American Scientist
    Because of his influence on countless readers, Gardner became known as the father of “recreational mathematics”—playful mathematical problems designed and ...<|control11|><|separator|>
  6. [6]
  7. [7]
    A closer look at the probabilities of the notorious three prisoners
    The “problem of three prisoners”, a counterintuitive teaser, is analyzed. It is representative of a class of probability puzzles where the correct solution ...Missing: explanation | Show results with:explanation
  8. [8]
    [PDF] Fundamentals of Probability - HKU EEE
    Total probability and Bayes' rule. Example: The Three Prisoners problem. Let XA, XB, XC be the events of sentencing prisoners A, B, C, respectively. Without ...Missing: P_A G_B
  9. [9]
    [PDF] STA 611: Introduction to Mathematical Statistics - Stat@Duke
    2.3 Bayes' Theorem. SKIP 2.4 The Gambler's Ruin ... 2.1 The Definition of Conditional Probability. Three Prisoners Problem ... 2.3 Bayes' Theorem. Bayes ...
  10. [10]
    [PDF] Bi=[X=i]
    Bayes Theorem. Suppose that you are walking down the street and notice ... Three prisoners problem. Three prisoners, A, B and C, are in separate cells ...
  11. [11]
    [PDF] closer look at the robabilities of the - Gwern
    Falk, R., 1992. A closer look at the probabilities of the notorious three prisoners. Cognition, 43: 197-223. The “problem of three prisoners”, a ...
  12. [12]
    The Monty Hall Dilemma Donald Granberg - Thad A. Brown, 1995
    The dilemma is whether to stick with the initial hunch or switch to the remaining alternative. The correct but counterintuitive solution is to switch. In the ...<|control11|><|separator|>
  13. [13]
    Why Humans Fail in Solving the Monty Hall Dilemma: A Systematic ...
    Jun 1, 2018 · The problem is also known as the 'Three Doors Problem' and is mathematically equivalent to the 'Three Prisoners Problem' (e.g., Shimojo & ...Missing: formal | Show results with:formal<|control11|><|separator|>
  14. [14]
    Why Almost Everyone Gets the Monty Hall Probability Puzzle Wrong
    Jul 16, 2024 · They've found that some of the most common cognitive biases may be to blame, along with a core misunderstanding of how probability works. ... the ...
  15. [15]
    [PDF] The Psychology of the Monty Hall Problem - Cerge-Ei
    To test possible transfer effects of training, we used the three-prisoners problem and a four-door version of the Monty Hall problem called the two goats-two ...
  16. [16]
    (PDF) The Psychology of the Monty Hall Problem - ResearchGate
    Oct 9, 2025 · ... three prisoners problem do not differ from the corresponding ones in the Monty Hall problem (Falk,. 1992; Ichikawa & Takeychi, 1990). The ...
  17. [17]
    [PDF] Three Prisoners
    Students are asked to test which of the two values, 1/3 or 1/2, reflects prisoner A's updated chances of perishing at dawn. Let us attempt to resolve the issue ...<|control11|><|separator|>
  18. [18]
  19. [19]
    [PDF] Gardner's Two Children Problems and Variations - ERIC
    For 26 years Martin Gardner wrote a famous and very widely read math- ematical puzzle column in the magazine Scientific American. In 1959, he published what ...
  20. [20]
    Calcul Des Probabilites : J. Bertrand - Internet Archive
    Nov 10, 2006 · Calcul Des Probabilites. by: J. Bertrand. Publication date: 1907 ... PDF download · download 1 file · SINGLE PAGE PROCESSED TIFF ZIP download.
  21. [21]
  22. [22]
    [PDF] Self-locating belief and the Sleeping Beauty problem
    My aim is to pose a problem arising from the interaction between these two sorts of uncertainty, solve the problem, and draw two lessons from the solution. 1.
  23. [23]
    [PDF] University of Huddersfield Repository
    The Three Prisoners Problem Revisited: issues in the use of Bayesian networks for security applications. Original Citation. Pickering, Jonathan and Xu, Zhijie ...
  24. [24]
    Indifference, neutrality and informativeness: Generalizing the Three ...
    Aug 7, 2025 · Wechsler et al. [2] extended the Three Prisoners Paradox by considering N prisoners, k executions and m announcements. To extend the Siblings ...
  25. [25]
    [PDF] The Sensor Management Prisoners Dilemma - AMOS Conference
    In this paper, the problem is formalized as a Partially Observable Markov Decision Process (POMDP).Missing: machine | Show results with:machine
  26. [26]
    Quantum version of the Monty Hall problem | Phys. Rev. A
    Jun 14, 2002 · A version of the Monty Hall problem is presented where the players are permitted to select quantum strategies.
  27. [27]
    Good Things Always Come in Threes: Three Cards, Three Prisoners ...
    Aug 6, 2025 · PDF | I review two classic problems: the Monty Hall Paradox and the Prisoner's Dilemma and a third, Three Cards Problem.
  28. [28]
    The Equiprobability Bias from a Mathematical and Psychological ...
    Similarly to the three cards problem, in the three prisoners problem (Lindley, 1971; Shimojo & Ichikawa, 1989) people have the impression that after discarding ...