The most famous probability problems in history, and why the smartest people consistently get them wrong.
Once the host opens a door to reveal a goat, only two doors remain: yours and one other. It seems obvious that the car is equally likely to be behind either one. You already made a choice, the host gave you new information, and now you're back to a coin flip. Switching seems arbitrary, a 50/50 guess either way. Many people feel that switching even implies less confidence in your original choice, making staying the "stronger" decision psychologically. This intuition is held with remarkable conviction, even by trained mathematicians who encounter the problem for the first time.
Always switch. Switching wins the car exactly 2 out of 3 times. Staying wins only 1 out of 3. The host's action did not make this a 50/50 problem, it revealed information that was always asymmetric. The key insight: if your initial door was wrong (which it is 2/3 of the time), the host is forced to reveal the only remaining goat, leaving the car behind the door you didn't pick. Switching captures all of those cases.
Let the car be behind Door 1, 2, or 3 with equal probability. You pick Door 1. The host opens Door 3 (a goat). Should you switch to Door 2?
Step 1, Prior probabilities:
$$P(\text{Car} = 1) = P(\text{Car} = 2) = P(\text{Car} = 3) = \tfrac{1}{3}$$Step 2, Host behavior (conditional probabilities):
If the car is at Door 1, the host can open Door 2 or Door 3 (both are goats). He chooses randomly:
$$P(\text{Opens 3} \mid \text{Car} = 1) = \tfrac{1}{2}$$If the car is at Door 2, the host must open Door 3 (only non-car, non-chosen door):
$$P(\text{Opens 3} \mid \text{Car} = 2) = 1$$If the car is at Door 3, the host cannot open Door 3. So:
$$P(\text{Opens 3} \mid \text{Car} = 3) = 0$$Step 3, Total probability of host opening Door 3:
$$P(\text{Opens 3}) = \tfrac{1}{3}\cdot\tfrac{1}{2} + \tfrac{1}{3}\cdot 1 + \tfrac{1}{3}\cdot 0 = \tfrac{1}{6} + \tfrac{1}{3} = \tfrac{1}{2}$$Step 4, Apply Bayes' Theorem:
$$P(\text{Car} = 1 \mid \text{Opens 3}) = \frac{P(\text{Opens 3}\mid\text{Car}=1)\cdot P(\text{Car}=1)}{P(\text{Opens 3})} = \frac{\tfrac{1}{2}\cdot\tfrac{1}{3}}{\tfrac{1}{2}} = \boxed{\frac{1}{3}}$$ $$P(\text{Car} = 2 \mid \text{Opens 3}) = \frac{1\cdot\tfrac{1}{3}}{\tfrac{1}{2}} = \boxed{\frac{2}{3}}$$Conclusion: After the host reveals Door 3, staying with Door 1 wins with probability 1/3. Switching to Door 2 wins with probability 2/3. The intuitive 50/50 answer is simply wrong, the host's knowledge and constrained behavior break the symmetry. The 2/3 probability assigned to "not your door" at the start does not split when one door is opened; it concentrates entirely on the remaining unchosen door.
There are 365 days in a year. Surely you'd need close to half that, about 183 people, before two of them are likely to share a birthday. Even 50 people feels like it should be nowhere near enough. Most people, when asked to estimate the group size needed for a 50% chance, guess somewhere between 100 and 200. The number 23 sounds impossibly small.
With just 23 people in a room, the probability that at least two share a birthday exceeds 50%. With 57 people, the probability passes 99%. The intuition fails because we naturally think about the probability that someone shares a birthday with us specifically, which is indeed low. But the problem asks about any pair from among all possible pairs, and 23 people create $\binom{23}{2} = 253$ distinct pairs.
It is far easier to compute the probability of the complementary event, no shared birthdays, then subtract from 1.
P(no shared birthday among n people):
Person 1 can have any birthday: $\frac{365}{365}$. Person 2 must avoid Person 1's birthday: $\frac{364}{365}$. Person 3 must avoid both: $\frac{363}{365}$. Continuing:
$$P(\text{all different}) = \frac{365}{365} \cdot \frac{364}{365} \cdot \frac{363}{365} \cdots \frac{365 - n + 1}{365} = \frac{365!}{(365-n)!\cdot 365^n}$$Therefore:
$$P(\text{at least one shared birthday}) = 1 - \prod_{k=0}^{n-1}\frac{365-k}{365}$$Key values (computed):
| n (people) | P(at least one match) |
|---|---|
| 10 | 11.7% |
| 20 | 41.1% |
| 23 | 50.7% |
| 30 | 70.6% |
| 40 | 89.1% |
| 57 | 99.0% |
| 70 | 99.9% |
Why intuition fails, counting pairs:
$$\binom{n}{2} = \frac{n(n-1)}{2}$$With $n = 23$: there are $\binom{23}{2} = 253$ distinct pairs, each one is an independent chance for a collision. The calculation above is essentially tracking the growing impossibility of 253 independent non-collisions all succeeding simultaneously.
The cryptographic birthday attack generalizes this: to find a collision in a hash function with $N$ possible outputs, you need roughly $\sqrt{N}$ random inputs, not $N/2$. For SHA-256 ($N = 2^{256}$), this reduces work from $2^{256}$ to "only" $2^{128}$, still enormous, but the principle is identical to the birthday problem.
A fair coin produces each sequence of three flips with equal probability: HHH, HHT, HTH, HTT, THH, THT, TTH, TTT, each with probability 1/8. If we each pick a different three-flip sequence and flip a coin repeatedly until one sequence appears, surely the game is fair? Every sequence is equally probable to appear on any given set of three flips, so both players should win 50% of the time. The game seems perfectly symmetric.
The game is deeply asymmetric. For any sequence your opponent names, you can always find a different sequence that appears before theirs with probability greater than 50%. The relationship between sequences is non-transitive: sequence A can beat B, B can beat C, and C can beat A, like a probabilistic rock-paper-scissors. Letting your opponent pick first is actually an advantage for the second player, not the first.
The non-transitive cycle β all 8 length-3 sequences:
| Opponent picks | You pick | Your winning probability |
|---|---|---|
| HHH | THH | 7/8 = 87.5% |
| HHT | THH | 3/4 = 75.0% |
| HTH | HHT | 2/3 = 66.7% |
| HTT | HHT | 2/3 = 66.7% |
| THH | TTH | 2/3 = 66.7% |
| THT | TTH | 2/3 = 66.7% |
| TTH | HTT | 3/4 = 75.0% |
| TTT | HTT | 7/8 = 87.5% |
Verified example β THH beats HHH with probability 7/8:
The only way HHH can appear before THH is if the sequence opens with three consecutive Heads. Any Tail before that would produce a THH first. Therefore:
$$P(\text{HHH appears first}) = P(\text{first 3 flips all Heads}) = \frac{1}{8}$$ $$P(\text{THH wins}) = 1 - \frac{1}{8} = \frac{7}{8}$$Conway's rule β how to choose your counter-sequence:
Given the opponent's three-flip sequence, your winning counter is constructed as follows: take the opponent's second element, reverse it (H becomes T, T becomes H), and place it in front of the opponent's first two elements.
| Opponent picks | Reverse 2nd element | Your counter | Your edge |
|---|---|---|---|
| HHT | H becomes T | T + HH = THH | 75% |
| THH | H becomes T | T + TH = TTH | 67% |
| HHH | H becomes T | T + HH = THH | 87.5% |
| TTT | T becomes H | H + TT = HTT | 87.5% |
For any sequence your opponent names, this rule produces a counter that wins more than half the time. Going first is the disadvantage, not the advantage.
With 100 passengers and a chaotic first boarder who randomly takes any seat, the subsequent boarding chaos seems impossible to analyze. Perhaps the last passenger's odds depend on everything that happened before, who sat where, which seats got displaced in sequence. Surely the answer changes with the number of passengers, and surely it's not something clean like exactly 50%. Most people either can't estimate it or guess something like 1/100 or some small number.
The last passenger gets their own seat with probability exactly 1/2, every single time, regardless of whether there are 2 or 2,000 passengers. This is one of the most elegant exact results in elementary probability. The last passenger either finds their seat empty (and takes it) or finds it taken (and takes some other seat), with equal probability.
Proof by mathematical induction.
Key observation: At any point during boarding, if Passenger 1's original seat is still empty, subsequent passengers will always take their own seat and the crisis "passes forward." The process only terminates in one of two ways: either Passenger 1's seat is taken first, or Seat 100's seat (the last passenger's seat) is taken first.
Base case ($n = 2$):
Passenger 1 randomly takes Seat 1 (probability 1/2) or Seat 2 (probability 1/2). If Seat 1: Passenger 2 takes their own seat. If Seat 2: Passenger 2 cannot take their seat. $P(\text{Passenger 2 gets own seat}) = 1/2$. β
Inductive step: Assume the result holds for $n-1$ passengers. For $n$ passengers, Passenger 1 takes:
- Seat 1 (prob $1/n$): Everyone else boards normally. $P(\text{last gets seat}) = 1$.
- Seat n (prob $1/n$): Last passenger's seat is taken. $P(\text{last gets seat}) = 0$.
- Seat $k$ for $2 \leq k \leq n-1$ (prob $(n-2)/n$): Passengers 2 through $k-1$ take their own seats, then Passenger $k$ faces the same problem with $n - k + 1$ remaining passengers. By the inductive hypothesis applied symmetrically, $P(\text{last gets seat}) = 1/2$.
Total probability:
$$P = \frac{1}{n}\cdot 1 + \frac{1}{n}\cdot 0 + \frac{n-2}{n}\cdot\frac{1}{2} = \frac{1}{n} + \frac{n-2}{2n} = \frac{2 + n - 2}{2n} = \frac{n}{2n} = \boxed{\frac{1}{2}}$$The proof works because at every step, Passenger 1's seat and the last passenger's seat are always equally likely to be chosen when a displaced passenger must pick randomly, creating a perfect symmetry that is preserved all the way to the end of the boarding process.
Why our pattern-seeking brains manufacture signal out of noise.
A basketball player has hit six shots in a row. They're clearly "hot", something has shifted in their form, concentration, or rhythm that makes continued success more likely. Every coach, commentator, and experienced fan knows to feed the ball to the hot hand. The streak is real, visible, and meaningful. Random sequences don't produce runs of six in a row, there must be an explanatory cause.
Random sequences always produce runs that look like streaks. In 100 fair coin flips, you should expect to see at least one run of 7 or more heads approximately 86% of the time. Our brains are streak-detectors optimized for a world where patterns usually mean something, so we see signal in noise. The statistical reality is more subtle: Gilovich et al. found no effect; Miller & Sanjurjo's 2018 correction suggests the data is consistent with a small genuine effect. The hot hand question remains scientifically contested.
Testing for independence: the key statistical question.
Define a player's shooting as a sequence of outcomes: $X_1, X_2, \ldots, X_n$ where $X_i = 1$ (hit) or $0$ (miss). The hot hand hypothesis claims:
$$P(X_{i+1} = 1 \mid X_i = 1) > P(X_{i+1} = 1 \mid X_i = 0)$$If shots are independent with constant success probability $p$, both conditional probabilities equal $p$.
Gilovich et al.'s finding (1985): For Philadelphia 76ers player data, the conditional probabilities were statistically indistinguishable. E.g., one player's data showed:
$$\hat{P}(\text{hit} \mid \text{prev. hit}) \approx 0.51, \quad \hat{P}(\text{hit} \mid \text{prev. miss}) \approx 0.54$$- actually slightly lower after a hit, consistent with defenders adjusting to a perceived hot player.
The MillerβSanjurjo bias (2018): In a finite sequence of $n$ flips with success probability $p$, if you condition on flip $k$ being a success and compute the proportion of successes among the immediately following flips within the same sequence, the expected proportion is not $p$, it is biased below $p$. For $n = 4$ and $p = 0.5$, the expected proportion of heads following a head is approximately $0.40$, not $0.50$.
The formula for the exact bias is complex, but intuitively: sequences with more late-sequence successes contribute fewer "following" observations, creating a systematic downward weight. When Gilovich et al.'s data is corrected for this bias, the corrected estimates are consistent with a positive hot hand effect of roughly 2β4 percentage points.
Expected length of longest run in $n$ flips of a fair coin:
$$E[\text{longest run}] \approx \log_2(n) - \log_2(\log_2(n)) + O(1)$$For $n = 100$: expected longest run $\approx \log_2(100) \approx 6.6$, so a run of 6 or 7 is statistically expected in any 100-trial sequence, not evidence of anything special.
After 10 consecutive reds on a roulette wheel, black feels overdue. The long-run statistics show roughly equal outcomes, and that can only happen if the universe "corrects" the imbalance eventually. A coin that has landed heads 10 times in a row must be more likely to land tails next, the very idea of probability implies eventual balance. This feels like common sense grounded in real statistical knowledge.
A roulette wheel and a fair coin have no memory. The probability of red on the next spin is always 18/38 on an American wheel (approximately 47.4%), regardless of what came before. The law of large numbers guarantees that proportions converge in the long run, but it says absolutely nothing about short-run "corrections." The wheel achieves long-run balance through sheer volume of independent trials, not through compensatory outcomes.
Formal definition of independence:
Events $A$ and $B$ are independent if and only if:
$$P(A \cap B) = P(A) \cdot P(B) \quad \Longleftrightarrow \quad P(A \mid B) = P(A)$$For successive roulette spins $R_1, R_2, \ldots, R_n$ (each Red with probability $p = 18/38$), statistical independence means:
$$P(R_{n+1} = \text{Red} \mid R_1 = \cdots = R_n = \text{Black}) = P(R_{n+1} = \text{Red}) = \frac{18}{38} \approx 0.474$$The Monte Carlo event quantified:
$$P(26 \text{ consecutive blacks}) = \left(\frac{20}{38}\right)^{26} \approx \frac{1}{136{\,}823{\,}184}$$This probability is indeed tiny. But given that this event happened:
$$P(27\text{th \,spin = Black} \mid \text{first 26 = Black}) = \frac{20}{38} \approx 0.526$$- exactly the same as on any spin. The rarity of the preceding sequence provides zero information about the next outcome.
What the Law of Large Numbers actually says:
$$\frac{1}{n}\sum_{i=1}^{n} X_i \xrightarrow{P} p \quad \text{as } n \to \infty$$The sample mean converges to $p$, but convergence happens because later outcomes are overwhelmed by the volume of all outcomes, not because later outcomes "compensate" for earlier ones. After 26 blacks, the expected total excess of blacks after $n$ more spins remains exactly 26 blacks (it doesn't shrink to zero). The fallacy confuses proportional convergence with numerical correction.
When events cluster in space or time, cancer cases in a neighborhood, crimes on a particular street, equipment failures in a short window, our brains immediately search for a cause. Uniform random scattering should look, well, uniform. Clusters mean something is driving events to that location. This intuition serves us well in many contexts: if every apple falling from a tree lands in one corner of the orchard, something is pulling them there. The same logic applied to human-scale data seems entirely reasonable.
Truly random (Poisson) processes always produce apparent clusters. The expectation of randomness being "spread out evenly" is fundamentally wrong, uniform distribution and random distribution are not the same thing. A perfectly uniform grid of bomb hits would actually be strong evidence that the Germans were deliberately targeting a pattern. Clarke's analysis showed the V-1 data matched the Poisson model with striking precision, the apparent clustering was entirely expected under randomness.
The Poisson distribution models the number of events in a fixed region when events occur randomly and independently at a constant average rate $\lambda$:
$$P(X = k) = \frac{e^{-\lambda}\lambda^k}{k!}, \quad k = 0, 1, 2, \ldots$$Clarke's London data (1946):
537 V-1 bombs fell across 576 equal squares. Expected hits per square: $\lambda = 537/576 = 0.9323$.
| Hits per square | Predicted (Poisson) | Actual count |
|---|---|---|
| 0 | $576 \cdot e^{-0.9323} \approx 226.7$ | 229 |
| 1 | $576 \cdot 0.9323 e^{-0.9323} \approx 211.4$ | 211 |
| 2 | $576 \cdot \frac{(0.9323)^2}{2}e^{-0.9323} \approx 98.5$ | 93 |
| 3 | $\approx 30.6$ | 35 |
| 4+ | $\approx 8.7$ | 8 |
Chi-squared goodness of fit:
$$\chi^2 = \sum_k \frac{(O_k - E_k)^2}{E_k}$$Clarke computed $\chi^2 \approx 1.17$ on 4 degrees of freedom. The $p$-value is approximately 0.88, meaning the observed data is more consistent with the Poisson model than 88% of Poisson-generated datasets would be. The fit is almost suspiciously good. There is no evidence whatsoever of deliberate targeting or unexplained clustering. The human perception of non-randomness in the data was entirely a cognitive artifact.
An athlete performs exceptionally well and earns the Sports Illustrated cover. The next season, their performance declines, clearly the attention, the pressure, the distractions, or perhaps the jinx itself caused the slump. Similarly, if you praise a student for an exceptional test score, they often do worse next time, suggesting praise creates complacency. If you scold someone for a terrible performance, they usually improve, suggesting criticism works. These patterns seem to have an obvious causal explanation.
Any extreme performance, elite or dismal, contains two components: skill and luck. Luck, by definition, does not persist. The next performance consists of the same skill component plus a new, independent luck component. Unless the new luck is also extreme in the same direction (improbable), the outcome will be closer to the mean. No jinx, no causal mechanism, no feedback loop required. The tragedy is that we attribute the mathematical rebound to our interventions, praise, or curses.
Galton's regression formula. Let $X$ be a parent's height (standardized: mean 0, SD 1) and $Y$ be a child's height (same scale). Galton found that the best linear predictor of $Y$ is:
$$\hat{Y} = r \cdot X$$where $r$ is the correlation coefficient between parent and child heights ($0 \leq r \leq 1$). Galton measured $r \approx 0.65$ for human height.
What this means: A parent who is 2 standard deviations above the mean is expected to have a child who is $r \times 2 = 0.65 \times 2 = 1.3$ standard deviations above the mean, noticeably closer to average.
Generalized regression formula:
$$\hat{Y} - \mu_Y = r \cdot \frac{\sigma_Y}{\sigma_X}(X - \mu_X)$$The regression fallacy explained: Suppose student true ability is $\mu$ and test score $= \mu + \epsilon$ where $\epsilon \sim N(0, \sigma^2)$ is random noise. A student who scored exceptionally high ($X = \mu + 3\sigma$) was likely helped by extreme positive noise. Their next test score:
$$E[Y \mid X = \mu + 3\sigma] = \mu + r \cdot 3\sigma \quad (< X \text{ when } r < 1)$$The expected decline toward $\mu$ happens even with zero intervention. Praising them appeared to "cause" a decline; scolding a low scorer appeared to "cause" improvement, but both observations are pure regression to the mean. As Kahneman writes, we are "rewarded" for punishing and "punished" for rewarding, creating systematically wrong beliefs about what works.
Four puzzles taken directly from our source book that overturn common assumptions about chance.
If Player A gets to pull the trigger first with a loaded 1-in-6 revolver, this is an overwhelming advantage. A gets a 1/6 chance of winning immediately before B even gets a turn. If A misses, B then gets their shot. The first-mover advantage seems large, perhaps a 60-70% edge. Many people estimate A wins 60-65% of the time, or even higher.
Player A wins with probability exactly $6/11 \approx 54.5\text{\%}$. Player B wins with probability $5/11 \approx 45.5\text{\%}$. The first-mover advantage is real but remarkably modest, barely more than a coin flip. The expected number of trigger pulls before the gun fires is exactly 6.
Setup: Players A and B alternate pulling the trigger of a 6-shot revolver with 1 bullet. A goes first. Each pull independently has probability $p = 1/6$ of firing.
Computing P(A wins):
A wins on turn 1 with probability $1/6$. A wins on turn 2 (A's second pull, after both missed once) with probability $(5/6)^2 \cdot (1/6)$. In general, A wins on their $k$-th pull with probability $(5/6)^{2k-2} \cdot (1/6)$:
$$P(\text{A wins}) = \frac{1}{6} + \frac{1}{6}\left(\frac{5}{6}\right)^2 + \frac{1}{6}\left(\frac{5}{6}\right)^4 + \cdots = \frac{1/6}{1 - (5/6)^2}$$Let $r = (5/6)^2 = 25/36$, so $1 - r = 11/36$:
$$P(\text{A wins}) = \frac{1/6}{11/36} = \frac{1}{6} \times \frac{36}{11} = \frac{6}{11} \approx 54.5\%$$ $$P(\text{B wins}) = 1 - \frac{6}{11} = \frac{5}{11} \approx 45.5\%$$Expected number of pulls before the gun fires:
$$E[\text{pulls}] = \frac{1}{p} = \frac{1}{1/6} = 6$$General formula for $n$ chambers, 1 bullet:
$$P(\text{A wins}) = \frac{n}{2n - 1}$$| Chambers | P(A wins) | P(B wins) |
|---|---|---|
| 2 | 2/3 = 66.7% | 1/3 = 33.3% |
| 6 | 6/11 = 54.5% | 5/11 = 45.5% |
| 10 | 10/19 = 52.6% | 9/19 = 47.4% |
| 100 | 100/199 = 50.3% | 99/199 = 49.7% |
As the number of chambers grows, the first-mover advantage shrinks toward zero. With six chambers the edge is real but modest.
With 10,000 uniquely numbered balls in a bag, drawing balls with replacement, you'd expect to need around 5,000 draws before you pull a repeated number. After all, at the midpoint you've drawn half the total population, so matches should start becoming likely. Most people anchor on the ratio of draws to population size and estimate somewhere between 2,000 and 5,000 draws.
A repeat appears with 50% probability after only about 118 draws, just 1.18% of the way through the population of 10,000. The intuition is off by a factor of roughly 42. This is the Birthday Problem with $N = 10{\,}000$ instead of 365. The key formula involves the square root of the population size, not half of it.
Mapping to the Birthday Problem: Drawing with replacement from 10,000 balls is identical to asking: how many people must be in a room before two share a birthday, if there are 10,000 days in the year?
Probability of no collision after $k$ draws:
$$P(\text{no repeat in }k\text{ draws}) = \frac{N}{N}\cdot\frac{N-1}{N}\cdot\frac{N-2}{N}\cdots\frac{N-k+1}{N} = \prod_{i=0}^{k-1}\frac{N-i}{N}$$Approximation for large $N$: Using $\ln(1-x) \approx -x$ for small $x$:
$$\ln P(\text{no repeat}) \approx -\sum_{i=0}^{k-1}\frac{i}{N} = -\frac{k(k-1)}{2N}$$ $$P(\text{no repeat in } k \text{ draws}) \approx e^{-k(k-1)/2N} \approx e^{-k^2/2N}$$Finding the 50% collision threshold:
$$P(\text{collision by draw }k) = 1 - e^{-k^2/2N} = 0.5$$ $$e^{-k^2/2N} = 0.5 \implies \frac{k^2}{2N} = \ln 2 \implies k = \sqrt{2N\ln 2}$$For $N = 10{\,}000$:
$$k^* = \sqrt{2 \times 10{\,}000 \times 0.6931} = \sqrt{13{\,}863} \approx \boxed{117.7 \approx 118}$$The square root structure: The collision threshold scales as $\sqrt{N}$, not $N/2$. This is the fundamental reason intuition fails, we expect linear scaling, but the math gives square-root scaling. For any population of size $N$:
$$k^*_{50\%} \approx 1.177\sqrt{N}$$This is why breaking an $n$-bit hash function requires approximately $2^{n/2}$ trials, not $2^n$, the birthday paradox directly dictates the security level of every hash-based cryptographic system.
With four people each flipping a fair coin, reaching a round where exactly one person gets a different outcome than the other three should take quite a few rounds. The chance that exactly one of four people gets tails (while others get heads) or exactly one gets heads (while others get tails) seems quite low, perhaps 10β15% per round, implying 7β10 rounds on average. For six people, the probability seems even lower, so more rounds would be needed.
For $n$ people, the probability of an odd-man-out result on any given round is $n/2^{n-1}$. For 4 people, this is $4/8 = 1/2$, giving an expected number of rounds of just 2. For 6 people, it's $6/32 = 3/16$, giving an expectation of about 5.3 rounds. For 3 people, expectation is $4/3 \approx 1.33$ rounds, faster than most expect.
Computing the per-round probability of an odd-man-out result:
An "odd man out" result occurs when exactly one person gets a different outcome from all others. With $n$ people each flipping a fair coin, this requires either exactly 1 Head among $n$ flips, or exactly 1 Tail among $n$ flips:
$$P(\text{odd man out in one round}) = \frac{\binom{n}{1} + \binom{n}{1}}{2^n} = \frac{2n}{2^n} = \frac{n}{2^{n-1}}$$Expected number of rounds (geometric distribution with success probability $p$):
$$E[\text{rounds}] = \frac{1}{p} = \frac{2^{n-1}}{n}$$| $n$ (people) | $P(\text{odd man out per round})$ | $E[\text{rounds}]$ |
|---|---|---|
| 2 | $2/2 = 1$ | 1 (always resolves immediately) |
| 3 | $3/4 = 0.750$ | $4/3 \approx 1.33$ |
| 4 | $4/8 = 0.500$ | $2.00$ |
| 5 | $5/16 = 0.3125$ | $3.20$ |
| 6 | $6/32 = 0.1875$ | $5.33$ |
| 8 | $8/128 = 0.0625$ | $16.0$ |
| 10 | $10/512 \approx 0.0195$ | $51.2$ |
The non-monotone intuition trap: Notice that 3 people ($E = 1.33$) resolves faster than 4 people ($E = 2$) even though 4 people have more "identifying information." This surprises most people. The expected rounds grow superexponentially: doubling the group from 5 to 10 increases expected rounds from 3.2 to 51.2, a 16Γ increase for a 2Γ increase in group size. This makes coin-flip odd-man-out impractical for groups larger than 6.
In a World Series between a dominant team (60% chance of winning each game) and an underdog (40%), the better team should win the series the vast majority of the time. Seven games should be more than enough to separate skill from luck. Most people estimate the favorite wins 75-85% of the time, or possibly even more.
The favorite wins the series approximately 71% of the time, meaning the underdog wins about 29% of series even when they are the clearly inferior team. Nearly 1 in 3 World Series should be won by the weaker team when the true per-game probability split is 60/40. To reduce the underdog's winning chance to below 10% would require a best-of-23 series.
Best-of-7 series: The first team to win 4 games takes the series. The series can last 4, 5, 6, or 7 games. Let $p = 0.4$ be the underdog's per-game win probability, $q = 0.6$ for the favorite. Define the underdog as Team A.
P(A wins series in exactly $k$ games) = P(A wins game $k$) Γ P(A won exactly 3 of the first $k-1$ games):
$$P(\text{A wins in }k\text{ games}) = \binom{k-1}{3}p^4 q^{k-4}, \quad k = 4, 5, 6, 7$$| Series length | Formula | Value ($p=0.4$) |
|---|---|---|
| 4 games | $\binom{3}{3}(0.4)^4(0.6)^0$ | 0.0256 |
| 5 games | $\binom{4}{3}(0.4)^4(0.6)^1$ | 0.0614 |
| 6 games | $\binom{5}{3}(0.4)^4(0.6)^2$ | 0.0922 |
| 7 games | $\binom{6}{3}(0.4)^4(0.6)^3$ | 0.1106 |
How many games to reduce the upset probability below 10%?
For a best-of-$n$ series (first to $(n+1)/2$ wins), the upset probability decreases slowly. Numerical computation shows:
| Series format | Upset probability ($p=0.4$) |
|---|---|
| Best-of-1 | 40.0% |
| Best-of-3 | 35.2% |
| Best-of-5 | 31.7% |
| Best-of-7 | 29.0% |
| Best-of-11 | 26.0% |
| Best-of-23 | $\approx 21\text{\%}$ |
The diminishing returns are stark: going from best-of-7 to best-of-23 eliminates fewer than 8 percentage points of upset probability. Skill and luck cannot be cleanly separated in any reasonably-lengthed playoff series.
Cases where more information makes our intuitive reasoning worse, not better.
Linda is described as: 31 years old, single, outspoken, very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations. Which is more probable: (A) Linda is a bank teller. (B) Linda is a bank teller and is active in the feminist movement? Option B feels more likely, it fits Linda's description much better. A bank teller who is also a feminist is a much more coherent, complete picture of who Linda is. Option A, by itself, seems oddly incomplete.
Option A is always at least as likely as Option B. This is a fundamental axiom of probability. "Bank teller AND feminist" is a subset of "bank teller." Any possible world where Option B is true is also a world where Option A is true, but not vice versa. Adding conditions can never increase probability. The vivid narrative made Option B feel more probable, but probability is not vividness.
The conjunction rule, a mathematical axiom:
For any two events $A$ and $B$ in any probability space:
$$P(A \cap B) \leq P(A) \quad \text{and} \quad P(A \cap B) \leq P(B)$$Proof (one line): $A \cap B \subseteq A$ (the intersection is a subset), and probability is monotone: if $X \subseteq Y$ then $P(X) \leq P(Y)$. Therefore $P(A \cap B) \leq P(A)$. This holds regardless of any information about Linda, any prior information, or any correlation between A and B. It is unconditional.
Visualized with a concrete partition:
$$P(\text{bank teller}) = P(\text{bank teller, feminist}) + P(\text{bank teller, not feminist})$$Since $P(\text{bank teller, not feminist}) \geq 0$:
$$P(\text{bank teller}) \geq P(\text{bank teller, feminist})$$Quantifying the representativeness heuristic: Kahneman and Tversky found that when subjects were asked to rank a list of 8 options, the conjunction was ranked higher than the constituent event by approximately 85% of participants. The fallacy is robust to education level, training in statistics, and explicit reminders about logic. It is driven by the System 1 heuristic of representativeness: how well does this outcome match the description?
The base rate for bank tellers: If 5% of people are bank tellers and 30% of people are feminists, with correlation $\rho = 0.1$:
$$P(\text{bank teller AND feminist}) = P(\text{bank teller}) \cdot P(\text{feminist} \mid \text{bank teller}) < P(\text{bank teller}) = 5\%$$No correlation structure can reverse this inequality. The conjunction is always bounded above by the less probable constituent event.
A disease test has 99% sensitivity (correctly identifies 99% of sick patients) and 99% specificity (correctly identifies 99% of healthy patients). Your test comes back positive. You reason: the test is 99% accurate. Therefore I almost certainly have the disease, the probability must be around 99%. This seems like straightforward application of the stated accuracy figure.
If the disease affects 1 in 1,000 people in the population, a positive result from a 99%-accurate test means only about a 9% chance you actually have the disease. How rare the disease is, the base rate, matters as much as how accurate the test is. A highly accurate test applied to a rare disease produces mostly false positives simply because there are so many more healthy people than sick people.
Bayes' Theorem, the fundamental tool:
$$P(\text{Disease} \mid \text{Positive}) = \frac{P(\text{Positive} \mid \text{Disease}) \cdot P(\text{Disease})}{P(\text{Positive})}$$Given values:
- $P(\text{Positive} \mid \text{Disease}) = 0.99$ (sensitivity / true positive rate)
- $P(\text{Positive} \mid \text{No Disease}) = 0.01$ (false positive rate = 1 β specificity)
- $P(\text{Disease}) = 0.001$ (1 in 1,000 base rate)
- $P(\text{No Disease}) = 0.999$
Expanding the denominator by total probability:
$$P(\text{Positive}) = P(\text{Pos}\mid\text{Dis})\cdot P(\text{Dis}) + P(\text{Pos}\mid\text{No Dis})\cdot P(\text{No Dis})$$ $$= 0.99 \times 0.001 + 0.01 \times 0.999 = 0.00099 + 0.00999 = 0.01098$$Applying Bayes:
$$P(\text{Disease} \mid \text{Positive}) = \frac{0.00099}{0.01098} \approx \boxed{0.0902 \approx 9\%}$$The natural frequency version (Gigerenzen's recommended framing): Imagine 100,000 people tested. Approximately 100 have the disease; 99 of them test positive. Of the 99,900 healthy people, approximately 999 also test positive (false positives). Total positives: $99 + 999 = 1{\,}098$. Of these, only 99 actually have the disease: $99/1{\,}098 \approx 9\text{\%}$.
The base rate dominance: How the answer changes with base rate, holding test accuracy constant at 99%/99%:
| Disease prevalence | P(Disease | Positive) |
|---|---|
| 1 in 10,000 | β 1.0% |
| 1 in 1,000 | β 9.0% |
| 1 in 100 | β 50.0% |
| 1 in 10 | β 91.7% |
For rare diseases, even extremely accurate tests produce a majority of false positives in the positive group. The implication for medicine: a positive screening test for a rare cancer is not a diagnosis, it is a prompt for a confirmatory test with higher specificity.
If Player A has a higher batting average than Player B in April AND in May AND in June, then over the full season, Player A must have a higher batting average than Player B. This is so intuitively obvious that it seems like a tautology. If A is better in every subgroup, A is better overall, how could it possibly be otherwise?
It can be otherwise, when the subgroup sizes are very different. The overall batting average is a weighted average of subgroup averages, where the weights are the number of at-bats. If Player A faced more pitchers in difficult conditions (during which both players hit poorly) while Player B faced more pitchers in favorable conditions, the weighted average can reverse the direction of every subgroup comparison. This is not a statistical error or a paradox in the logical sense, it is a mathematical fact about weighted averages.
The JeterβJustice example (1995β1997 MLB seasons, a famous documented instance):
| Season | Derek Jeter | David Justice |
|---|---|---|
| 1995 | 12/48 = .250 | 104/411 = .253 β |
| 1996 | 183/582 = .314 | 45/140 = .321 β |
| 1997 | 190/654 = .291 | 163/495 = .329 β |
| Combined | 385/1284 = .300 β | 312/1046 = .298 |
Justice outperformed Jeter in every individual season. Yet over the three seasons combined, Jeter's overall average (.300) exceeds Justice's (.298). The reversal occurs because Jeter accumulated far more at-bats in seasons where both players hit well, and fewer at-bats in the single season where both hit their worst, the weighting tilts his composite average upward.
The mathematical structure:
Let $a_1/b_1$ and $a_2/b_2$ be player A's rates in two groups, and $c_1/d_1$ and $c_2/d_2$ be player B's rates. Suppose $a_1/b_1 < c_1/d_1$ and $a_2/b_2 < c_2/d_2$. It is still possible that:
$$(a_1 + a_2)/(b_1 + b_2) > (c_1 + c_2)/(d_1 + d_2)$$The paradox arises when $b_1/b_2 \neq d_1/d_2$, when the two players allocate their "sample sizes" differently across the two groups. If A concentrates more sample in the group where both players perform best, A's composite can exceed B's despite trailing in every subgroup.
Formal condition for Simpson's Paradox:
$$\frac{a_1}{b_1} < \frac{c_1}{d_1}, \quad \frac{a_2}{b_2} < \frac{c_2}{d_2}, \quad \text{yet} \quad \frac{a_1+a_2}{b_1+b_2} > \frac{c_1+c_2}{d_1+d_2}$$This requires the "mixing weights" $b_1/b_2$ and $d_1/d_2$ to differ in a specific direction. When this condition holds, no statistical analysis of aggregate data can recover the true subgroup relationships without explicit subgroup separation.
Someone tells you: "I have two children and at least one is a boy." Since one child is a boy, the other child is either a boy or a girl, a coin flip. Therefore the probability that both are boys is 1/2. If instead they said "my older child is a boy," again the younger is either a boy or a girl, still 1/2. The two statements seem to provide the same information, so they should yield the same answer.
The two statements give different answers because they convey information through different mechanisms. "At least one is a boy" eliminates one of four equally likely birth-order combinations, leaving three. The answer is 1/3. "The older child is a boy" eliminates exactly two combinations (older is a girl), leaving two equally likely ones. The answer is 1/2. The information content is genuinely different even though it sounds almost identical.
The four equally likely possibilities for two children (assuming equal probability of boy/girl and independent births):
| Older child | Younger child | Label |
|---|---|---|
| Boy | Boy | BB |
| Boy | Girl | BG |
| Girl | Boy | GB |
| Girl | Girl | GG |
Version 1: "At least one child is a boy."
This eliminates only GG. Remaining sample space: {BB, BG, GB}, each equally likely.
$$P(\text{both boys} \mid \text{at least one boy}) = \frac{1}{3}$$Version 2: "My older child is a boy."
This eliminates GB and GG. Remaining sample space: {BB, BG}, each equally likely.
$$P(\text{both boys} \mid \text{older is boy}) = \frac{1}{2}$$The Tuesday Boy Variant (Foshee, 2010): "I have two children. One is a boy born on a Tuesday." What is the probability both are boys?
Expand the sample space to include 7 days per sex: $7 \times 2 = 14$ outcomes per child, $14^2 = 196$ equally likely pairs. Outcomes with "at least one boy born on Tuesday": $13 + 14 - 1 = 27$ (by inclusion-exclusion, where 13 are cases where the older child is a Tuesday-boy and the younger can be anything, plus 14 where the younger is a Tuesday-boy, minus 1 overlap).
Of these 27, how many have both boys? Both boy outcomes where at least one is a Tuesday-boy: $7 + 7 - 1 = 13$. Therefore:
$$P(\text{both boys} \mid \text{one is a Tuesday boy}) = \frac{13}{27} \approx 0.481$$Not 1/3, not 1/2, but almost 1/2. The day of birth genuinely changes the probability, even though it seems utterly irrelevant. The paradox shows that in conditional probability, how a piece of information is obtained is part of the mathematical problem, not just background context.
Problems where the mathematically optimal strategy feels absurd, and yet provably works.
When hiring for a position with $n$ applicants interviewed in random order, most people adopt one of two naive strategies: decide early (perhaps after 5-10 candidates) to avoid missing the best candidates, or interview everyone and then decide (but this violates the "can't go back" constraint). Neither extreme seems clearly optimal, and the right cutoff seems highly dependent on the total number of candidates. The problem feels too situation-specific to have a clean mathematical solution.
There is an exact optimal strategy: reject the first $\lfloor n/e floor$ candidates unconditionally (using them only to set a "benchmark"), then immediately hire the first subsequent candidate who beats all previously seen candidates. As $n o \infty$, the optimal cutoff fraction converges to $1/e \approx 36.8\%$ and the probability of selecting the best candidate also converges to $1/e \approx 36.8\%$.
Formal setup: $n$ candidates appear in uniformly random order. You observe each one and must immediately accept or reject with no recall. Goal: maximize the probability of selecting the single best candidate.
The threshold strategy: Reject the first $r - 1$ candidates unconditionally (use them to build a benchmark). Then accept the first subsequent candidate who beats everyone seen so far.
Probability of success with cutoff $r$:
$$P(\text{success}) = \frac{r-1}{n} \sum_{k=r}^{n} \frac{1}{k-1}$$Finding the optimal cutoff as $n$ grows large: Set $t = r/n$ (fraction rejected). The probability becomes:
$$P(\text{success}) \approx -t \ln t$$Maximize by setting the derivative to zero: $d(-t \ln t)/dt = -\ln t - 1 = 0$, giving $t^* = 1/e \approx 0.368$.
$$P(\text{success at optimum}) = \frac{1}{e} \approx 36.8\%$$Results for specific pool sizes:
| Pool size | Candidates to skip | Success probability |
|---|---|---|
| 5 | 1 | 43.3% |
| 10 | 3 | 39.9% |
| 20 | 7 | 38.4% |
| 50 | 18 | 37.4% |
| 100 | 36 | 37.1% |
| 1,000 | 367 | 36.8% |
Even with a perfect strategy you find the best candidate only about 37% of the time. No strategy can do better. The 63% failure rate is the irreducible cost of sequential, no-recall decision-making under uncertainty.
There are two envelopes, one containing twice as much money as the other. You open yours and find $100. You reason: the other envelope has either $50 (if yours is the larger) or $200 (if yours is the smaller), each with probability 1/2. Expected value of the other envelope: $(1/2)(50) + (1/2)(200) = 125$. Since $125 > 100$, you should switch. But after switching, you can apply the same logic again, and again. You should keep switching forever, generating infinite free money from nothing. Something is wrong.
The flaw is subtle but definitive: after opening and observing $x$ (say, 100 dollars), you cannot simultaneously assign equal probability to "mine is the smaller ($x = a$)" and "mine is the larger ($x = 2a$)" without implicitly assuming a probability distribution over the possible values of the smaller amount $a$, and no such distribution exists that is both coherent and makes the expected-value calculation valid. The variable $x$ is not independent of which envelope you hold, so treating $x$ as fixed while assigning probabilities to both cases is a form of double-counting.
The seductive argument:
Let $a$ be the smaller amount (unknown). You open your envelope and find $x$. Either $x = a$ (yours is smaller) or $x = 2a$ (yours is larger), each with probability 1/2:
$$E[\text{other envelope}] = \frac{1}{2}(2x) + \frac{1}{2}\left(\frac{x}{2}\right) = \frac{5x}{4} > x$$This says you should always switch, regardless of $x$, even before opening the envelope.
The flaw, the two cases are not symmetric:
When you observe $x$, you're conditioning on the event "my envelope contains $x$." The two cases are:
- Case 1: $a = x$, so other envelope $= 2x$
- Case 2: $a = x/2$, so other envelope $= x/2$
These cases have probability $P(a = x)$ and $P(a = x/2)$ respectively. Setting both equal to 1/2 requires:
$$P(a = x) = P(a = x/2) \quad \text{for all } x > 0$$This would require a distribution over $a$ such that every value is equally likely to be $a$ or $a/2$. But this means $P(a = t) = P(a = t/2) = P(a = t/4) = \cdots$ for all $t$, which forces either $P(a = t) = 0$ for all $t$ (not a distribution) or $P(a = t) = \infty$ (also not a distribution). No proper probability distribution over $a$ can make the two cases equally likely for all observed values of $x$.
The correct analysis:
Suppose $a$ is known to be drawn from a proper distribution $f(a)$. Then after observing $x$:
$$P(\text{Case 1} \mid x) = \frac{f(x)}{f(x) + f(x/2)/2} \cdot \frac{1}{1}, \quad P(\text{Case 2} \mid x) = \frac{f(x/2)/2}{f(x) + f(x/2)/2}$$(Details omitted for brevity.) For any proper prior $f$, this produces a well-defined posterior, and the expected value of switching can be either positive or negative depending on $f$ and $x$. The paradox evaporates once you commit to a specific prior, but no "ignorance prior" (uniform over all positive reals) exists as a proper distribution.
If buses arrive on average every 10 minutes, a passenger arriving at a random time should wait about 5 minutes on average, halfway through the average gap. The logic seems airtight: on average you arrive in the middle of a gap of average length 10, so you wait half of that.
You almost always wait more than 5 minutes. The key error is assuming that a random arrival is equally likely to fall in any gap. But longer gaps occupy more time, so a random arrival is more likely to fall in a long gap than a short one. This length-biased sampling effect means the gap you're in at arrival is longer on average than the typical gap, and your expected wait exceeds half the typical inter-arrival time.
Setup: Buses arrive at intervals $T_1, T_2, T_3, \ldots$ where each $T_i$ is drawn i.i.d. from a distribution with mean $\mu = E[T]$ and variance $\sigma^2 = \text{Var}(T)$. A passenger arrives at a uniformly random time.
The length-biased distribution: The probability that your random arrival falls in a gap of length $t$ is proportional to $t$ (longer gaps are "larger targets"):
$$f_{\text{gap you're in}}(t) = \frac{t \cdot f(t)}{\mu} \quad \text{(length-biased distribution)}$$Expected length of the gap you're in:
$$E[T_{\text{gap}}] = \frac{E[T^2]}{\mu} = \frac{\mu^2 + \sigma^2}{\mu} = \mu + \frac{\sigma^2}{\mu}$$This always exceeds $\mu$ when $\sigma^2 > 0$ (i.e., whenever gaps are not all identical).
Expected waiting time:
Given you land uniformly within a gap of length $L$, your expected wait is $L/2$. Averaging over the length-biased distribution:
$$E[\text{wait}] = \frac{E[T^2]}{2\mu} = \frac{\mu}{2} + \frac{\sigma^2}{2\mu}$$The wait exceeds $\mu/2$ by exactly $\sigma^2/(2\mu)$, proportional to the variance of inter-arrival times.
Concrete example: Buses arrive either every 5 minutes or every 15 minutes, each equally likely.
$$\mu = \frac{5+15}{2} = 10, \quad E[T^2] = \frac{5^2 + 15^2}{2} = \frac{25+225}{2} = 125, \quad \sigma^2 = 125 - 100 = 25$$ $$E[\text{wait}] = \frac{125}{20} = 6.25 \text{ minutes}$$Despite the 10-minute average gap, you wait 6.25 minutes on average, 25% longer than the naive 5-minute estimate. Moreover, you're in a 15-minute gap 75% of the time (since 15/20 of the "covered time" is in 15-minute gaps), even though only 50% of gaps are 15 minutes long.
The friendship paradox (same mathematics): On average, a person's friends have more friends than that person does. Proof: you are more likely to be friends with popular people (they appear in more "friendships"), exactly analogous to arriving in a longer gap.
Some lottery numbers just seem more likely than others. Sequential combinations like 1-2-3-4-5-6 feel less likely, surely the lottery doesn't produce perfectly ordered sequences. Numbers that have "come up recently" feel more likely. Numbers that haven't appeared in a while feel overdue. And picking a "nice" pattern seems wasteful, real random sequences don't look patterned. Conversely, picking birthdays (1β31) feels smart since those numbers feel familiar and "lucky.".
Every combination of 6 numbers has exactly the same probability of being drawn: $1/\binom{49}{6} = 1/13{\,}983{\,}816$ in a standard 6/49 lottery. Sequence 1-2-3-4-5-6 is exactly as likely as any other combination. The real strategic consideration is entirely different: you want to pick combinations that other players avoid, so that if you win, you don't share the jackpot. Maximizing your expected jackpot share, not your probability of winning, is the tractable optimization problem.
The fundamental combinatorial fact:
In a standard 6/49 lottery, the total number of equally likely combinations is:
$$\binom{49}{6} = \frac{49 \times 48 \times 47 \times 46 \times 45 \times 44}{720} = 13{,}983{,}816$$Every single combination has exactly the same probability of winning:
$$P(\text{any specific combination wins}) = \frac{1}{13{,}983{,}816}$$This is identical for 1-2-3-4-5-6, for 7-14-21-28-35-42, and for every other combination. The draw has no memory and no preference.
The jackpot-splitting problem:
If $k$ other players chose the same combination and your ticket wins, your share is $1/(k+1)$ of the jackpot $J$. Your expected return per ticket is:
$$E[\text{return}] = \frac{1}{13{,}983{,}816} \times \frac{J}{k+1}$$The probability of winning is fixed. The only factor you can influence is $k$ β how many others picked the same numbers. Choosing less popular combinations lowers $k$ and raises your expected share.
Documented selection biases (UK National Lottery):
| Number type | Why people pick them | Sharing risk |
|---|---|---|
| 1 to 31 | Match birthdays and dates | High β heavily over-picked |
| Multiples of 7 | Perceived as lucky | Above average |
| 32 to 49 | No birthday link | Low β under-picked |
| 1, 2, 3, 4, 5, 6 | Chosen as a joke | Very high β thousands pick it |
| Diagonal patterns on slip | Visual convenience | Above average |
Selecting all six numbers above 31 does not improve your odds by a single decimal place. But it can reduce your expected number of co-winners by 20 to 30%, raising your expected payout if you do win. The only rational strategy in the lottery is not about which numbers get drawn. It is entirely about which numbers other people avoid.