Insight Editions · No. 19
The Drunkard's Walk
How Randomness Rules Our Lives
Leonard Mlodinow
Leonard Mlodinow traces the history of probability and explains how chance shapes everyday judgments, careers, markets, and other outcomes often attributed to skill.
A study guide to The Drunkard's Walk by Leonard Mlodinow. Shared free for personal reading; not for sale and not endorsed by the authors or publishers. Please read and buy the original.
Prologue
Mlodinow opens with a joke that is not really a joke. A Spanish lottery winner explains his ticket ending in 48: he dreamed of the number 7 for seven nights, "and 7 times 7 is 48." We laugh at the arithmetic, but we all filter the world through a theory of our own making, and our errors around chance are usually just as large, only less visible.
The prologue sets out the book's brief. Since the 1930s researchers have known that people can neither produce a random sequence nor recognise one. A whole field, judgment and decision making, grew from that, and won a Nobel Prize in Economics (Kahneman, 2002), yet its lessons have not reached ordinary life. The book aims to fix that by telling the history of the ideas of chance and showing them at work in politics, business, medicine, sport and law.
Three opening exhibits preview the argument. Coach firings in major sports have, on average, no effect on results. CEO cults misfire: Gary Wendt left GE Capital with a star reputation and a $45 million signing bonus to run Conseco; the stock tripled, then the company went bankrupt two years after he quit. The question is not whether he was good, but whether one past success predicts the next. And the economist Armen Alchian's phrase "fortuitous circumstances" names the thing we keep forgetting: success and failure can arise from neither great skill nor great incompetence.
A drunkard's walk (random walk) is the path of a molecule knocked about by its neighbours. Mlodinow uses it as a metaphor for careers, families and golf rounds, and promises that the same mathematics that describes the molecule helps us read our own lives.
| Ch | Tool | Human error it exposes |
|---|---|---|
| 1 | Coin models, regression | Mistaking luck for skill |
| 2 | Three laws of probability | Conjunction, availability, bad multiplication |
| 3 | Sample space | Mis-listed outcomes (Monty Hall) |
| 4 | Counting, expectation | Ignoring coincidence and rare costs |
| 5 | Law of large numbers | Law of small numbers, gambler's fallacy |
| 6 | Bayes | Inverted conditionals, base-rate neglect |
| 7 | Bell curve, margin of error | False precision in grades, wine, polls |
| 8 | Social statistics, regression | Expecting individuals to be predictable |
| 9 | Significance testing | Seeing patterns, illusion of control |
| 10 | Random walk, normal accidents | Hindsight; judging people by results |
The prologue's exhibits in full
Coach firings. Across major sports, mathematical analysis of coach dismissals finds no average effect on team performance. Teams fire after bad runs; the next season often looks better; credit goes to the new hire. Regression toward the mean predicts the bounce without any coaching miracle.
CEO cults. Gary Wendt left GE Capital under Jack Welch with a star reputation and a $45 million signing bonus to run troubled Conseco. The stock tripled within a year. Two years after he quit, Conseco was bankrupt and the stock traded for pennies. Kodak, Lucent and Xerox supply parallel cases. The open questions: was the task impossible, was he asleep, or were the assumptions wrong that an executive has near-absolute power and that one past success predicts the next?
Missing information. Confirming global warming took enormous effort because sparse data invite competing readings. Drugs are declared safe and later pulled. Doctors and patients misread effectiveness statistics. Parents and teachers misread SATs. Wine critics misread point scores. Investors misread five-year fund returns.
Armen Alchian named the blind spot: success and failure can arise from neither great skill nor great incompetence but from "fortuitous circumstances." The mind assigns each event a definite cause and resists unrelated or random factors. The first step is to see those factors at work.
The title borrows a mathematical term for random motion of molecules. The same tools that describe a molecule's path help read careers, families and golf rounds. The book's goal is to illustrate chance in everyday affairs and to leave the reader seeing life with a deeper understanding of randomness.
Reading map for the ten chapters
Chapters 1–2 name the human errors (stories over frequencies; conjunction and bad products). Chapters 3–4 supply the repair kit of sample spaces and counting. Chapters 5–6 add the law of large numbers and Bayes. Chapters 7–8 scale to measurement error and social aggregates. Chapters 9–10 return to pattern hunger, markets, war, accidents and careers. Keep Alchian's "fortuitous circumstances" in view the whole way: skill is real, and so is the remainder.
What "fortuitous circumstances" changes
Alchian's phrase is a permission slip to stop moralising every outcome. A coach firing that coincides with a bounce does not prove coaching skill. A CEO hired after a boom does not prove executive omnipotence. A drug declared safe on thin data is not proved safe. Chocolate milkshakes as heart-healthy satire in the prologue is Mlodinow's joke about motivated reading of evidence: we see the pattern we came to see. The book that follows is training in seeing the other pattern - the footprint of chance - without denying skill where it is real.
FORTUITOUS. Alchian's word for the blind spot: success and failure can arise without great skill or great incompetence. The drunkard's walk is the molecule metaphor for the same claim. If you can recite fortuitous circumstances, you have the prologue.
Pick one recent "destiny" story (a coach firing, a CEO hire, a product hit or flop). Make three columns: skill, chance, attempts. Fill one concrete fact per column. Keep the card; revisit it after Chapter 10.
Invert the prologue: suppose every outcome is skill. Then coach firings that coincide with bounces prove coaching genius, Conseco's arc proves executive omnipotence, and Alchian's phrase is an excuse. The rest of the book then has nothing to teach. Hold that challenge through Chapter 1.
Chapter 1. Peering through the Eyepiece of Randomness
As a teenager Mlodinow watched the yellow flame of the Sabbath candles and asked why his father, who did not believe in God, still prayed. The answer opened a war story he rarely told. In Buchenwald, starving, his father stole a loaf of bread from the bakery. The baker had him hauled before the Gestapo. A death sentence seemed certain. Instead he was assigned as the baker's assistant: the man who was to punish him needed help more than he needed an example. Chance, not virtue or planning, decided whether that day ended in death or in a job that kept him alive. The chapter's theme is already there: outcomes that look like destiny are often noise plus a story told afterward.
We use intuitive processes when we assess uncertain situations. Those processes evolved for a world of immediate threats and small samples, not for mutual funds, studio slates or medical tests. The mechanisms by which people analyse chance are a product of evolution, wiring and experience, and they are systematically biased. Making wise assessments under uncertainty is a rare skill; like any skill, it can be trained once the biases have names.
Brains that fight over probability
A laboratory game makes the bias measurable. Lights flash green 75 percent of the time and red 25 percent. Subjects guess the next colour. Two strategies dominate. Maximising always picks the more frequent colour and scores near 75 percent. Matching guesses green 75 percent of the time and red 25 percent, trying to reproduce the mix, and scores only 62.5 percent (0.75×0.75 + 0.25×0.25). Rats and other animals tend to maximise. Humans hunt for a pattern, match the frequencies, and lose. We beat ourselves by insisting the sequence must have a plot.
Split-brain research sharpens the point. After the corpus callosum is cut, the right hemisphere tends to maximise on such tasks; the left seeks patterns and invents explanations. The left-brain interpreter will fabricate a reason for an action it did not choose. Ordinary life runs the same script: a streak in sports, markets or hiring gets a story, and the story feels like insight even when the data are noise.
When one option is clearly more frequent, always picking it beats trying to reproduce the mix. Pattern hunting feels intelligent and lowers accuracy. The left-brain interpreter would rather be wrong with a story than right without one.
Bertrand Russell named the default stance naive realism: the doctrine that things are as they seem. We treat the visible scoreboard as the person. Randomness makes that doctrine expensive. Stockbrokers recommend funds that "did well"; teachers rewrite a pupil's character after one exam; voters credit or blame leaders for swings that weather and commodity prices already explain.
Kahneman, Tversky and regression toward the mean
Their heuristics-and-biases program, and later prospect theory, mapped how people misread uncertain evidence. A seed case was flight instructors. Instructors noticed that harsh criticism after a bad landing was often followed by improvement, and warm praise after a good landing by a worse one. They concluded that punishment works and praise fails. Kahneman saw regression toward the mean: in any series that mixes skill and luck, extreme performances tend to be followed by more ordinary ones. The instructor's theory fits every data point and is still wrong.
| Ingredient | Role |
|---|---|
| Skill | Stable part of performance |
| Luck | Transient noise, up or down |
| Extreme trial | Often skill plus unusual luck |
| Next trial | Luck re-samples; result moves toward mean |
| False lesson | "Punishment worked" / "Praise spoiled them" |
Kahneman wondered whether such misconceptions were universal. They were. The same fog covers SAT scores, wine ratings, mutual-fund rankings and medical tests. A single number feels decisive; the error around it stays invisible until you model it. Parents, teachers and students treat exam ranks as pure ability. Connoisseurs treat point scores as pure quality. Investors treat five-year returns as pure skill. Each is a measurement with noise.
Rejection as a measurement of noise
Suppose four publishers reject your thriller about love, war and global warming. Does that prove the book is bad? History says no. Between a finished work and a hit sits a gulf of randomness and uncertainty. Many books destined for great success had to survive repeated rejection:
- Anne Frank, The Diary of a Young Girl: rejected by a string of houses; later on the order of 30 million copies.
- John Grisham, A Time to Kill: rejected by about 26 publishers; The Firm later made him a phenomenon.
- Theodor Geisel (Dr. Seuss), And to Think That I Saw It on Mulberry Street: 27 rejections.
- J. K. Rowling, Harry Potter: about 9 London rejections before Bloomsbury; an editor's child liked the opening.
- John Kennedy Toole, A Confederacy of Dunces: rejected until after his suicide; his mother forced the manuscript on Walker Percy; it won a Pulitzer.
- Sylvia Plath collected early rejections; George Orwell's Animal Farm was turned down by T. S. Eliot at Faber as too political; Isaac Bashevis Singer waited years; Tony Hillerman was told to "get rid of all that Indian stuff."
William Goldman and David Picker's Hollywood maxim applies beyond film: nobody knows anything. A lot of what happens in careers, investments and life decisions, both good and bad, is less a clean mirror of ability than a draw from a noisy process.
Hollywood: anointment, regicide and pipeline lag
We habitually underestimate randomness. A broker recommends a fund that beat the market; a studio fires an executive after a cold streak; a board crowns a CEO after a boom. Hollywood makes the error expensive and visible.
Ability still matters: a jittery homemade horror video does not become a hit as easily as a polished thriller. But the residuals are huge. The Blair Witch Project cost about $60,000 and took in about $140 million. Ishtar cost about $55 million and returned about $14 million. Last Action Hero arrived with Arnold Schwarzenegger and a marketing blitz, then collapsed. American Graffiti cost under $1 million and earned about $115 million. Star Wars: United Artists and Universal passed; Fox's Alan Ladd Jr. backed it; George Lucas took a roughly $200,000 fee and kept merchandising; the film made about $461 million on a $13 million budget. Green-light decisions are made years before release; films are subject to weather, news cycles, competing openings and moods no executive controls.
- Blair Witch: ~$60k → ~$140M.
- Ishtar: ~$55M → ~$14M.
- American Graffiti: <$1M → ~$115M.
- Star Wars: ~$13M budget → ~$461M; prior studio passes.
- Star executive contracts still reach $25 million on the theory that past green lights predict future ones.
Coin models for "genius" streaks
Imagine 10 executives each toss a fair coin 10 times. The chance that at least one gets 8 or more results of one kind is about two-thirds. Boards will still crown that person. Or run a television test: two episode versions, A and B, each shown to 20,000 viewers. The most probable pattern is that one version leads throughout; a seesaw of lead changes is far rarer. Mlodinow notes it is about 88 times more likely that one leads the whole way than that the lead keeps flipping. A stable lead looks like proof of quality. It can be sampling noise. Such issues are rarely discussed in boardrooms, so the typical response to a cold streak is a ritual firing.
Sherry Lansing led Paramount through a run that magazines treated as personal genius: Forrest Gump, Braveheart, strong years in market share. Under her, Paramount's share ran roughly 11.4, 10.6, 11.3, then 7.4, 7.1, 6.7. Variety headlines that had celebrated a "banner year" now diagnosed failure. Commentators invented theories of her "losing her touch"; Timeline and Lara Croft: Tomb Raider became exhibits. She left; the next year share rebounded to nearly 10 percent, driven largely by films she had already approved. The pipeline lag matters: today's release is yesterday's green light.
Mark Canton at Columbia shows the reverse. Projects he left behind, including Men in Black (about $589 million worldwide) and other hits, became triumphs for successors. Michael Ovitz lasted fifteen months as Disney president and left with a package near $140 million. David Begelman forged checks at Columbia and was still treated as a player. Sumner Redstone, Brad Grey and others rotate through the same anointment-regicide cycle. Hollywood pays for a narrative of control that the variance cannot support.
:::key Anointment and regicide A short sample of hits creates a genius; a short sample of misses creates a failure. Both readings ignore regression, pipeline lag (films green-lit years earlier) and the base rate that someone will look hot when many people take many swings.
In mathematical terms Lansing's fate has a short and a long explanation. The short one: seven years is a small sample of studio outcomes. The long one: even a skilled executive sits on top of a process with enormous variance; good luck at the start and bad luck at the end need not imply a change in talent. Hindsight made her firing look inevitable. The same hindsight would have looked different if two releases had swapped years.
Roger Maris and the coin
Mlodinow's own epiphany came in a college mathematics course that taught him to see hidden randomness. The seed memory was 1961. He was barely of reading age, but he recalled the faces of Roger Maris and the more popular Mickey Mantle chasing Babe Ruth's 60 home runs. Mantle's knees failed; he finished at 54. Maris reached 61. Fans booed; sportswriters moralised; Commissioner Ford Frick floated an asterisk because the season was longer than Ruth's.
Years later, influenced by that course, Mlodinow modelled the season. From yearly statistics you can estimate a player's probability of a home run at each opportunity. For Maris that rate was about 1 per 14.7 plate appearances. To simulate a "normal Maris" you need a rather odd coin and a strong wrist, or a computer. Matching Ruth's 60 is about 1 season in 32 (~3 percent) for that player. History shows about one player every three years with both the talent and the playing time to be in that conversation; over 70 years the chance that someone matches Ruth by luck is a little over 50 percent. Mlodinow had expected the odds to resemble Jack Whittaker's lottery win; they were far less extreme. Stephen Jay Gould and physicist Ed Purcell pressed related points about streaks and chance in baseball.
| Piece | Figure |
|---|---|
| HR chance per opportunity | ~1 / 14.7 |
| Matching Ruth, one "normal Maris" | ~1 season in 32 |
| Comparable players | ~1 every 3 years |
| Someone matches over ~70 years | a little over 50% |
| Mantle 1961 | 54 HR (injury-limited) |
| Maris 1961 | 61 HR |
We can never know for certain whether Maris was a far better player in 1961 than in his other seasons. We can know that extraordinary accomplishments in sports, and elsewhere, are what you should expect from noisy processes with many trials. Extreme success is not automatic proof of a new kind of person. When we look at record years we should keep in mind that extremity is cheap when the sample is large.
What chapter 1 demands
Peering through the eyepiece of randomness means accepting that visible success is a blend. Ability matters; so do the number of attempts, the variance of the domain, pipeline delays and the stories we invent after we see the score. The rest of the book supplies the mathematics that makes that claim precise: sample spaces, counting, the law of large numbers, Bayes, the bell curve, social statistics, significance tests and the drunkard's walk itself.
Hot-hand research remains contested. Miller and Sanjurjo (2018, Econometrica) showed a streak-selection bias in classic tests and recovered a hot-hand effect after correction; their 2024 Review of Economics and Statistics paper reports robust hot-hand performance across four controlled shooting experiments and finds experts can identify who gets hot. A 2026 NBA permutation/hierarchical analysis, by contrast, finds little streak behaviour beyond chance once in-game covariates are controlled. Practical takeaway: controlled shooting can show small hot-hand effects, while many hiring, fund and coaching streak stories still collapse under regression and multiple-comparison controls. Short runs remain weak evidence of changing executive skill.
Further exhibits from the chapter
Naive realism (Russell) is the doctrine that things are what they seem. We start there and only slowly learn that perception and scoreboards are filtered. Kahneman's flight-instructor story was the seed that made him ask whether misconceptions about feedback were universal; the heuristics-and-biases program grew from that question.
Goldman / Picker. The Hollywood saying "nobody knows anything" is not nihilism about craft. It is a claim about prediction: green lights are made years before release, and the residual variance swamps most executive skill signals in short samples.
Episode A/B. When two versions of a show are tested on large equal audiences, the most probable path is that one leads throughout. A seesaw of lead changes is rare (about 88× less likely than a one-sided lead in Mlodinow's illustration). Stable leads feel like quality. They can be sampling structure.
Canton pipeline. Mark Canton's Columbia slate left behind hits (including Men in Black at roughly $589 million) that successors claimed. Pipeline lag makes succession narratives unreliable: today's hit may be yesterday's green light.
Gould and Purcell. Stephen Jay Gould and physicist Ed Purcell analysed baseball streaks as chance processes. Extreme seasons are expected when many players take many trials. Maris in 1961 fits that frame without requiring a supernatural peak.
Worked regression and coin models
Flight instructor arithmetic. Let true skill be stable and luck be symmetric. Conditioning on "landing in the bottom decile" selects for bad luck; the next trial's expected luck is near zero, so the expected landing rises even if teaching did nothing. Conditioning on the top decile selects for good luck; the next trial falls. Any feedback policy that punishes the bottom and praises the top will look empirically validated and still be causally backwards.
Ten executives, ten coins. P(one specific executive gets ≥8 heads) = C(10,8)+C(10,9)+C(10,10) over 2¹⁰ = (45+10+1)/1024 ≈ 5.5 percent. P(at least one of ten does) = 1-(1-0.055)¹⁰ ≈ 43 percent for heads alone; counting extremes either way lands near the book's two-thirds intuition that someone will look "hot." Boards that promote on a single hot sample are buying that tail.
Maris simulation sketch. Opportunity count × (1/14.7) gives a season mean; Poisson or binomial tails put 60 HR near the 1-in-32 season rarity for that mean. Multiply by the number of eligible player-seasons across decades and "someone reaches Ruth" becomes ordinary. Gould and Purcell's streak work is the same logic applied to hits in a row rather than homers in a year.
More Hollywood and baseball texture
Variety headlines tracked Paramount's share with the mood of a sports page: "Banner Year" on the way up, autopsy prose on the way down. Brad Grey and Sumner Redstone appear in the surrounding corporate cast as reminders that the anointment-regicide cycle outlives any one executive. Hostess Ding Dong product-placement colour in Mlodinow's Hollywood tour is comic relief with a point: the town's folklore is thick, its controlled experiments thin.
On the baseball side, Mickey Mantle's knees and Babe Ruth's 60 set the emotional frame that made Maris's 61 a moral referendum. Jack Whittaker's lottery win is Mlodinow's foil: he expected Maris-level rarity to match lottery-level rarity and found the home-run tail far fatter once the right coin model was in place. That calibration error - treating every extreme as maximally extreme - is the lay version of ignoring the sample space of player-seasons.
REGRESS. Extreme today, closer to average tomorrow. Hollywood anointments, coach firings and Maris-style seasons are samples in a noisy mix of ability and luck, not destinies.
After your next extreme result (a great review, a bad quarter, a hot streak), write one sentence on what regression toward the mean predicts for the next trial before you invent a cause. Compare that sentence to the story you would have told without it.
Invert Chapter 1: suppose short streaks prove that skill itself changed. Then fire after every dip, promote after every spike, and treat Maris as supernatural. The flight-instructor feedback loop becomes wisdom, and the rest of the book's maths is optional colour.
- Prefer maximising on clear base rates; stop inventing patterns that cut accuracy.
- After any extreme result, ask what regression toward the mean predicts for the next trial.
- Read Hollywood anointments, coach firings and record seasons as samples, not destinies.
- Treat rejection lists and flops of heavily backed projects as evidence that ex ante judgment is noisy.
2 The Laws of Truths and Half-Truths
Probability theory, Mlodinow says, is a codification of common sense that keeps producing uncommon results. Unlike geometry, born among philosophers, it grew from gambling and spells. This chapter gives three basic laws, shows how intuition breaks them, and follows the ideas from Greek bones to the California Supreme Court.
Linda the bank teller
Tversky and Kahneman described Linda: 31, single, outspoken, a philosophy major concerned with discrimination who joined antinuclear protests. Eighty-eight subjects ranked eight statements about her. Average ranks (1 = most likely): "feminist" 2.1, "bank teller and feminist" 4.1, plain "bank teller" 6.2. 85 percent ranked the conjunction above the single statement. With only three options, 87 percent still did. Even 36 graduate students reminded of the conjunction rule mostly corrected, but two held firm.
The probability that two events both occur can never exceed the probability of either one alone. P(A) = P(A and B) + P(A and not B). Every "and" can only shrink the odds.
The error vanishes when the details are irrelevant (no one thinks "owns a pancake franchise and had a sex change" beats either alone). It appears when the added detail fits the story. As Tversky and Kahneman put it, "a good story is often less probable than a less satisfactory explanation." 91 percent of internists judged a pulmonary embolism more likely to cause a rare symptom plus a common one than the rare symptom alone. Among 200 lawyers asked about the Paula Jones v. Clinton suit, those given specific routes to an early end (settlement, withdrawal, dismissal) rated an early end more likely than those asked the plain question. Specific paths inflate probability in the mind even as they cut it in the maths.
An evolutionary story: pouncing on a blur that turns out to be leaves costs less than ignoring a real lizard. Seeing a few mirages may be the price of seeing patterns at all. Treat this as plausible rather than proven. Frequency wording ("of 100 women like Linda, how many...") cuts the error sharply, as Gerd Gigerenzer argued.
Why the Greeks had no probability
The Greeks gambled with astragali, heel bones with four stable faces landing about 10, 10, 40 and 40 percent of the time. The prized Venus throw (four different faces) has probability about 384 in 10,000, but no Greek worked it out. Mlodinow gives four reasons:
- Belief that outcomes were the will of the gods (Herodotus and oracles; chance as divine speech).
- A philosophy that prized proof over likelihood (in Plato's Phaedo and Theaetetus, arguments from probabilities are "impostors").
- Poor frequency records: no culture of tabulating throws.
- A clumsy alphabetic number system without zero. The zero as a true number came with the Indian mathematician Mahavira in the ninth century.
The availability bias
Are there more six-letter words ending in ing or with n as the fifth letter? Most say ing, yet every -ing word has n fifth. We judge frequency by ease of recall, so vivid memories inflate. We feel cursed at the checkout (true odds of picking the slowest of five lines: about 1 in 5). In a mock drunk-driving trial, the side whose evidence was told vividly ("What do you want, I should paint 'em pink?") won, and more so after a 48-hour delay when the vivid detail was what remained.
The Romans and the half proof
| Law | Rule | Example |
|---|---|---|
| Conjunction | P(A and B) is at most P(A) | Linda |
| Product | If independent, P(A and B) = P(A) × P(B) | two no-shows: 1/3 × 1/3 = 11% |
| Sum | If mutually exclusive, P(A or B) = P(A) + P(B); all outcomes sum to 1 | both or neither: 44% + 11% = 55% |
Product and sum in practice
The product rule's escape clause is independence. A married police officer's chance of divorce (1 in 50) and death on the job (1 in 5,000) in one year multiply to 1 in 250,000, but slightly less in reality, since the dead cannot divorce. Dating examples make the same point: if the chance a stranger is both attractive to you and available is roughly 1/10 × 1/10, you need many trials, not one perfect evening. An airline's two late passengers each show with probability 2/3; if independent, both arrive with probability 4/9 ≈ 44%, neither with 1/9 ≈ 11%, exactly one with 4/9; if they travel together, the multiplication is simply wrong. Applied correctly, two independent half proofs give three quarters of a proof, and no finite pile of partial proofs ever reaches certainty. We no longer use half proofs, but we still misuse products.
Modern courts: DNA and the Collins couple
We no longer use half proofs, but we use one-in-a-billion proofs. Experts quote random-match odds of 1 in a billion, yet lab errors (swapped samples, mixtures, misreporting) run around 1 percent by many estimates and are often excluded from testimony. Jurors average the two and land near 1 in 500 million. The sum rule says the total error is about the larger term: roughly 1 in 100. Timothy Durham got over 3,100 years despite eleven alibi witnesses, because a lab failed to separate victim and rapist DNA; he served nearly four years before a retest freed him. Thomas Lyon at USC is among those who have pressed courts to face lab error as the dominant term.
A blonde with a ponytail snatched a purse and fled in a yellow car driven by a Black man with a beard and moustache. Witnesses could not identify the Collinses, so the prosecution called a maths instructor who multiplied six "probabilities" (yellow car 1/10, moustache 1/4, Black man with beard 1/10, ponytail 1/10, blonde 1/3, interracial couple in car 1/1,000) into 1 in 12 million and called that the chance of innocence. Two errors: the traits are not independent (beards come with moustaches; dropping that factor gives about 1 in a million), and the relevant number is not the chance a random couple matches but the chance a matching couple is guilty. In a population of several million there might be 2 or 3 matching couples, so guilt on this evidence was 1 in 2 or 3. The California Supreme Court reversed and condemned "trial by mathematics."
The "Linda problem" still replicates, though researchers debate how much is a logic error and how much comes from reading "bank teller" as "bank teller who is not a feminist"; frequency wording cuts the error sharply, as Gigerenzer argued. In forensics, probabilistic genotyping software now reports likelihood ratios for mixed DNA samples, and reviews such as the US PCAST report (2016) pushed courts to ask for measured lab error rates, the very number Mlodinow says juries were denied. Proficiency testing has improved but error rates for complex mixtures remain contested.
Extra detail: availability, dating and partial proofs
Availability is not only a word-count trick. News coverage makes plane crashes feel common and diabetes feel rare; personal memory makes your own slow checkout lane feel cursed. The pink-truck mock trial shows that vivid phrasing beats dry phrasing even when the facts are matched, and the gap grows after delay because vivid fragments are what survive.
Dating maths. If the chance a stranger meets your attractiveness bar is about 1 in 10, and the chance they are available and interested is another 1 in 10, independence gives about 1 in 100 per meeting. That is not a counsel of despair; it is a reason to increase trials rather than rewrite your personality after three quiet weekends.
Airline passengers. Two travellers each show with probability 2/3. If independent: both 4/9, neither 1/9, exactly one 4/9. If they are a couple moving together, those products do not apply. The product rule is a tool with a precondition, not a slogan.
Rome's additive habit overstated certainty; modern DNA rhetoric that quotes only the random-match probability repeats the habit under a new name.
Worked numbers for the three laws
Conjunction (Linda). If P(bank teller) were even as high as 0.05 and P(feminist | teller) were 0.01, the conjunction would be 0.0005, far below either margin. Ranking the conjunction above "teller" alone is not a small slip; it is an inversion of the partial order probability imposes.
Product (independence). Two fair dice show sixes with probability 1/36, not 1/6. Two independent medical false-negative rates of 10 percent leave a 1 percent chance both miss. Dependence changes everything: two symptoms of the same disease are not independent evidence.
Sum (exclusive). The chance of rolling a 7 with two dice is 6/36 because the six ordered pairs are exclusive and exhaustive among the seven-ways. Non-exclusive events need inclusion-exclusion: P(A or B) = P(A)+P(B)-P(A and B). Roman half proofs failed here by treating overlapping uncertainties as additive bricks.
DNA arithmetic. Random-match probability 10⁻⁹ beside lab-error probability 10⁻² yields a combined failure risk on the order of the lab term. Jurors who "average" the exponents invent a number that exists in no theory. Durham's 3,100-year sentence on unseparated mixed DNA is the cost of that invention.
Apply the three laws at work
Hiring. A CV that is "engineer and amateur novelist and ultramarathoner" is a conjunction: each "and" narrows the set. Prefer separate scores for skills you need over a vivid package that feels uniquely right.
Incident reviews. Two weak signals that are dependent (same sensor family) do not multiply into strong proof. Independent detectors do.
Legal memos. Quote lab error beside match probability. If you cannot get a proficiency-test rate, say so; do not let a 1-in-a-billion figure travel alone.
Product analytics. "Users who did A and B and C convert more" is often a conjunction of rarity with a story. Compare against the base rate of C alone.
More from Linda's eight statements and Roman practice
The full Linda battery asked subjects to rank eight statements, not three. Ranks for colourless options sat low; ranks for story-rich conjunctions sat high. The short three-option version still produced 87 percent conjunction errors. Graduate students who could recite the rule still slipped until the problem was reframed in frequencies. International House of Pancakes and sex-change filler items show the boundary: when the added conjunct is absurd rather than representative, the fallacy fades.
Cicero's attack on astrology after Cannae is an early base-rate argument: mass death does not imply a shared horoscope. Justinian's Digest graded proof; bishop elections demanded 72 witnesses. The error was algebraic (adding half proofs) rather than moral. Modern courts that admit match probabilities without lab-error rates repeat the algebraic error with better fonts.
- Every added detail lowers probability, even as it raises plausibility.
- Multiply only for independent events; add only for mutually exclusive ones.
- With two error sources, the larger one dominates: check the lab, not the headline odds.
- Ask "how likely is guilt, given a match?" not "how likely is a match?"
AND LOWERS. Every added "and" lowers probability while raising the story's pull. Linda beats the bank teller because the conjunction feels like a biography. Multiply only for independent events; with two error sources, the larger one dominates.
Find one news or work claim that joins two traits with "and" (person, product, or suspect). Write P(A), P(B), and P(A and B) under independence. Note how much the conjunction drops. Then ask whether the events are truly independent before you trust any product.
Invert Chapter 2: suppose vivid detail raises probability. Then Linda-style stories beat base rates, Roman half-proofs look modern, and courtroom matches become verdicts. The three laws become pedantry that loses to a good narrative.
The three laws are the arithmetic of everyday judgment. Chapter 3 asks the prior question: before you multiply or add, have you listed the right outcomes? Cardano's sample space, and Monty Hall's ten thousand angry letters, are next.
3 Finding Your Way through a Space of Possibilities
A ruined genius in Rome
In 1576 an oddly dressed old man shuffled through Rome, shouting at no one. Gerolamo Cardano had been Europe's famous physician and astrologer, chair of medicine at Pavia, inventor of a forerunner of the combination lock and the universal joint, author of 131 books. Before he died he burned 170 manuscripts; among the 111 survivors was a 32-chapter treatise, The Book on Games of Chance, the first book ever written on the theory of randomness. Its key idea Mlodinow calls the law of the sample space, a probability version of balancing a chequebook.
Monty Hall and the 10,000 letters
In September 1990 Marilyn vos Savant (listed by Guinness with an IQ of 228) answered a reader in Parade: three doors, one car, two goats. You pick a door; the host, who knows where the car is, opens another door to show a goat and offers a switch. Should you switch? She said yes. About 10,000 letters arrived, 92 percent saying she was wrong, nearly 1,000 from PhDs. "How many irate mathematicians are needed to change your mind?" wrote one. Paul Erdős, one of the century's great mathematicians, refused to believe the proof and conceded only after watching a computer simulation come out 2 to 1 for switching. Martin Gardner had posed a version in 1959, noting that "in no other branch of mathematics is it so easy for experts to blunder."
If a random process has many equally likely outcomes, the probability of a favourable outcome is the proportion of outcomes that are favourable. The set of all outcomes is the sample space. If the outcomes are not equally likely, weight each by its odds.
Getting the outcomes right
The subtle part is deciding what counts as an outcome. The eighteenth-century mathematician d'Alembert said two coin tosses give 0, 1 or 2 heads, so each has probability 1/3. Wrong: the true outcomes are the sequences HH, HT, TH, TT, so one head is 50 percent and the others 25 percent each. A bookmaker offering 3 to 1 on one head would be a gift.
Relabel heads and tails as girls and boys and you get the two-daughter problem, an isomorphism: one problem in disguise as another. Fraternal twins: chance of at least one girl is 75 percent. Now the twist: given that one child is a girl, what is the chance both are? Not 1/2. Strike BB from the sample space, leaving GB, BG, GG, so 1 in 3. If instead you are told the first child is a girl, strike BB and BG, and it is 1 in 2. Which information you were given changes the space.
| Question | Naive answer | Correct |
|---|---|---|
| 1 head in 2 tosses | 1/3 | 1/2 |
| At least one girl, twins | 1/2 | 3/4 |
| Both girls, given one is | 1/2 | 1/3 |
| Two liars name the same flat tyre | 1/16 | 1/4 |
The Duke flat-tyre story: two students skip a final, claim a flat, and are asked separately "which tire?" Vos Savant's reader thought matching by chance was 1 in 16. It is 1 in 4: whatever the first says, the second has a 1 in 4 chance of matching.
Solving Monty Hall
Pick door 1. The car is behind door 1, 2 or 3, each with probability 1/3.
- Lucky Guess (probability 1/3): the car is yours; the host opens either other door at random. Switching loses.
- Wrong Guess (probability 2/3): the host must open the one goat door you did not pick, leaving the car behind the other. Switching wins.
The host is not random: he uses what he knows, and that knowledge flows to the remaining door. Switching wins 2 times in 3, and statistics from the show agree. With 100 doors, your first pick has a 1 in 100 chance; the host opens 98 goat doors; the one remaining door carries 99 in 100.
Cardano's life: chance in the flesh
Mlodinow tells Cardano's life as a run of luck good and bad. His mother tried to abort him; he was born apparently lifeless and revived in warm wine; he caught bubonic plague as a baby in 1501 and survived when his nurse and half-brothers died. His father, a geometer and sometime friend of Leonardo, used the boy's head as a reading desk.
He worked in an age before the equals sign (Robert Recorde, 1557) and before x for multiplication, with Hindu-Arabic place value and fractions only recently arrived. Barred by the Milan College of Physicians, he moved to Piove di Sacco, where people preferred sorcerers, and with his empty diary he wrote his gambling book. A lucky recovery by a prominent prior made him famous; he became rector of the College that had refused him, published the Ars magna (with the first clear treatment of negative numbers), and chaired medicine at Pavia.
Then his children ruined him. His son Giovanni poisoned his wife and was executed; enemies had Cardano exiled; Tartaglia, furious that Cardano had published his method for cubic equations, helped persuade his other son Aldo, a torturer for the Inquisition, to inform on him. His gambling book appeared in 1663, by which time others had surpassed it.
Mlodinow leans on Cardano's own autobiography and hostile contemporaries, so some lurid details are contested by historians. The mathematics is the solid part: Cardano's book handled equally likely outcomes well but made no systematic count of sequences, the gap Pascal would fill.
Monty Hall is now a stock test for AI reasoning systems. Large language models answer the textbook version correctly but often fail variants (a host who opens a door at random, or doors that are transparent), giving the memorised "always switch" answer where it no longer applies. Benchmark designers in 2024 to 2026 use exactly such perturbed puzzles to separate reasoning from recall. The lesson is Cardano's: define the sample space from the actual rules, not from a familiar story.
Extra detail: Ask Marilyn, Erdős, twins and tires
Twins and conditioning. "At least one girl" among two children (assuming BG/GB/GG/BB equally likely, excluding BB) leaves three cases, so two girls is 1/3. "The first child is a girl" leaves the other child at 1/2. "At least one girl born on Tuesday" enlarges the space and raises the chance of two girls above 1/3. The words of the condition change the space.
Duke flat tires. Two couples, two cars, a night ending with two flats looks like 1/16 if you naively square 1/4. Once you fix which tire failed on which car under the story's constraints, the coherent probability is 1/4. Mis-listing is the crime; arithmetic is the accessory.
d'Alembert listed outcomes of "at least one head in two tosses" as H, T, HT and got 2/3 instead of 3/4. Cardano's generation still lacked a clean algebra of sequences; Pascal later closed that gap. Martin Gardner's 1959 column made the three-door puzzle a public object; vos Savant made it a national argument.
Worked sample spaces
Monty Hall table. Car behind 1, 2 or 3 with probability 1/3 each; you always start at door 1; host opens the lowest-number goat door available. Stay wins only in the car-at-1 row (1/3). Switch wins in the other two rows (2/3). Any intuition that says "two doors left, hence half" has already left the sample space.
Three-prisoners / 100-doors. The same asymmetry scales. With 100 doors, stay wins with 1/100; switch to the host's leftover door wins with 99/100. The host's policy is the information channel.
Cardano's gap. He counted faces well and sequences poorly. That is why d'Alembert-style errors persisted until Pascal and Fermat's letters of 1654 made ordered outcomes standard. The Book on Games of Chance (1663 print) is a beginning, not an ending.
Apply sample space thinking
A/B tests with peeking. The sample space includes every look you took at the dashboard. Optional stopping without correction inflates false wins.
Support tickets. "At least one critical bug in the release" is not the same condition as "the first reported bug is critical." Write the condition before you estimate.
Security. An attacker who knows which doors are empty (Monty) is not the same as a random scanner. Model the adversary's policy or you will compute the wrong switch odds.
More from the Monty Hall storm and Cardano's century
A Gallup poll after the vos Savant column found that a large majority of Americans sided against her; the puzzle had become a referendum on common sense. On the television show, host Monty Hall and model Carol Merrill (once "Miss Azusa") performed the reveal as entertainment, not as a probability seminar. Robert Jarvik and other public figures weighed in; Guinness World Records listed vos Savant's IQ. None of that settles the maths. The settled maths is the sample-space table: stay 1/3, switch 2/3, provided the host always opens a goat door and never opens the car.
Cardano titled his manuscript in ways that later editors rendered as The Book on Games of Chance. He distributed copies thinly; the 1663 print was posthumous. He lived through the arrival of Hindu-Arabic place value in European practice, the equals sign (Recorde, 1557), and the cubic wars with Tartaglia. His autobiography and enemies' pamphlets disagree on lurid details; the durable residue is the insistence that you must count the ways before you claim an edge. That is the same demand Monty Hall makes of anyone who answers "50-50."
Full Monty Hall decision tree in words
Label doors 1, 2, 3; you always pick 1; the host always opens a goat door, choosing at random if both remaining doors have goats.
- Car at 1 (prob 1/3): host opens 2 or 3 each with prob 1/2. Switch to the remaining door → lose.
- Car at 2 (prob 1/3): host must open 3. Switch → win.
- Car at 3 (prob 1/3): host must open 2. Switch → win.
Summing, switch wins with probability 1/3 + 1/3 = 2/3. If you prefer conditional probability: after host opens door 3, P(car at 1 | open 3) = 1/3, P(car at 2 | open 3) = 2/3. The same arithmetic that shocked Ask Marilyn's mailbag is three lines once the host's policy is written down. Variants that break the policy (host opens at random, or can open the car) change the answer; that is why AI benchmarks now perturb the puzzle.
LIST FIRST. Wrong list, wrong odds. Monty Hall is not a coin flip: the host knows, so switching wins with probability 2/3. Conditioning removes points from the sample space; ask what was ruled out, and how.
On paper, label doors 1, 2, 3; always pick 1; have the host open a goat door. Enumerate the three car locations and mark when switching wins. Do not stop until the 2/3 result appears from the list, not from a remembered slogan.
Invert Chapter 3: treat the host as random. Then switching is 50-50, Erdős was right to resist, and sample-space discipline is theatre. Any informed actor (seller, insider, algorithm that "reveals") gets misread the same way.
- List the outcomes before you count them; most errors are in the list.
- Conditioning removes points from the sample space: ask what was ruled out, and how.
- An informed actor (a host, a seller, an insider) is not random; their choices carry information.
4 Tracking the Pathways to Success
Cardano's work made no splash because his world preferred incantation to calculation. Superstition was ordinary: merchants stole a hanged man's clothes to boost beer sales (and a bond trader in 2003 confessed to CNN that using the "wrong" toilet brought bad luck). Cardano himself thought losing streaks came because "fortune is averse." What changed was the scientific revolution.
Galileo and the grand duke's dice
In 1583, legend says, a Pisa student timed a swinging cathedral lamp against his pulse and found the period independent of the swing's width. Galileo's method, experiment plus mathematics, also produced a grudging consulting note, "Thoughts about Dice Games," for the grand duke of Tuscany, who had noticed that 10 beats 9 with three dice, by about 8 percent.
Both totals can be made from six combinations, so why the difference? Because combinations are not equally likely: (6,3,1) can arrive in 6 orders, (3,3,3) in only one. Counting ordered outcomes gives 27 ways to roll 10 and 25 to roll 9, a ratio of 1.08.
The chance of an event depends on the number of ways it can occur. The idea is obvious; the size of the effect, and the difficulty of counting, are not.
On a 10-question true-false quiz, a guesser can score 100 percent in 1 way, 90 percent in 10 ways and 80 percent in 45 ways. In a class of 25 pure guessers, there is about a 75 percent chance someone gets a B or better.
Coincidences that are not
Canadian lottery officials drew 500 car winners from 2.4 million subscriber numbers without crossing numbers off; one man won two cars. The odds of a repeat were about 5 percent, which is the birthday problem: with 23 people, a shared birthday is more likely than not, because what matters is the number of pairs (253). Germany's Lotto 6/49 repeated a winning set in 1995 (first drawn 1986), the first repeat in 3,016 draws; the chance of some repeat by then was about 28 percent.
Pascal, de Méré and the problem of points
Blaise Pascal, born 1623, joined the Mersenne circle (Descartes, Fermat) at 13. Chronically ill, he was told by doctors to stop working and amuse himself; his inheritance in 1651 began a "worldly period." Through it he met the gambler chevalier de Méré, who in 1654 asked the problem of points: if a game between equals is interrupted, how should the pot be split? Pascal began a famous correspondence with Pierre de Fermat, a Toulouse judge and the greatest amateur mathematician in history (who rose by outliving superiors felled by plague, and had himself been wrongly reported dead).
Mlodinow's example is the 1996 World Series. The Braves led the Yankees 2-0. Assume each game is a coin toss and imagine all five remaining games are played: 32 sequences. The Yankees need 4 of the 5, which happens in 5 + 1 = 6 ways, so their chance was 6 in 32, about 19 percent. (Counting games that would never be played is legitimate: the sequences exist whether or not anyone plays them.) The Yankees won the next four.
| True edge | Weaker side wins a best-of-7 | Series length for 95% confidence |
|---|---|---|
| 55-45 | about 4 in 10 | best of 269 |
| 2/3-1/3 | about 1 in 5 | best of 23 |
| Employee 60-40, best of 5 | weaker wins almost 1/3 | decades of comparisons |
It is dangerous to judge ability by short-term results. A "world champion" may simply be the luckier of two near-equals, and the same holds for quarterly rankings of firms or staff.
Pascal's triangle
For bigger counts you need a formula. The triangle where each number is the sum of the two above it was known to Jia Xian around 1050 and Zhu Shijie in 1303; Cardano mentioned it in 1570; Pascal put it to work and got the credit ("one of us places the ball better"). Row n lists the number of ways to choose 0, 1, 2... items from n. Row 5 (1 5 10 10 5 1) gives the Yankees' counts directly; row 10 gives the quiz counts. Choosing a table of 10 from 100 wedding guests has so many possibilities that checking one per second would take about 10,000 billion years.
Six people, a product that truly appeals to half the market. Row 6 is 1 6 15 20 15 6 1. Only 20 of 64 splits are an accurate 3-3; 44 of 64, about two thirds, show a lopsided "consensus" that is pure noise. A 5-to-1 verdict from six people is not evidence.
The night of fire and the invention of expectation
On 23 November 1654, weeks after the last letter to Fermat, Pascal had a two-hour religious trance, sewn into his coat lining as a note found after his death. He gave away his possessions, wore a spiked belt, and renounced geometry. Yet in the Pensées he applied probability to faith. Pascal's wager: if God exists, piety pays infinitely; if not, it costs a little. Multiply each payoff by its probability and add: the result is infinite. The method is mathematical expectation, often called the founding move of decision theory and game theory.
Expected value = sum of (probability x payoff). Mlodinow's 25-cent parking meter really costs $2.25, because he gets a $40 ticket about 1 time in 20. Three $400 mirror strikes in 2,400 reversals make his driveway a 50-cent toll road. A free $5 million sweepstakes with 1 in 200 million odds is worth 2.5 cents an entry, less than the stamp; the post office would have netted about $80 million.
He describes a state lottery as a game in which one player gets a fortune and, on NHTSA figures, about one player dies driving to buy a ticket. States set prizes so that expected winnings are below the ticket price.
Pick 6 of 44 gives 7,059,052 combinations. The pot, with lower prizes, was $27.9 million, so each $1 ticket was worth about $3.95. Adjusting for the risk of sharing (based on 170 past draws: no other winner 120 times), the expected value was still $23.4 million, or $3.31 a ticket. Australian organisers raised about $3,000 each from 2,500 investors, filled 1.4 million slips by hand, and worked 125 outlets, but bought only about 5 million tickets before the deadline. They held the winner anyway, and after a month of legal wrangling Virginia paid.
Pascal died in August 1662, aged 39, after the usual bleedings and purges.
Virginia and other US lotteries closed the loophole long ago, but the pattern recurs: in the 2010s, groups exploited roll-down rules in Massachusetts' Cash WinFall. The birthday problem now matters in security, where hash "collisions" appear after roughly the square root of the number of possible values, the reason modern systems use 256-bit hashes and why post-quantum cryptography standards (NIST, finalised 2024) size keys with these counts in mind. Sports analysts also accept Mlodinow's point: MLB's expanded playoffs since 2022 have produced several low-seeded champions, as short-series maths predicts.
Extra detail: Galileo, Pascal's night, focus groups and expectation
Galileo and three dice: faces summing to 10 have 27 microstates; summing to 9 have 25. The grand duke's intuition that they should be equal was the sample-space error of an age that still counted combinations more readily than ordered outcomes.
Pascal's night of fire (23 November 1654) and the Pensées wager belong to the same mind that split pots with Fermat. Expectation is the bridge: value each outcome by probability × payoff, then sum. Parking: a 25¢ risk of a ticket plus (1/20)×$40 expected fine equals $2.25, which loses to a 50¢ driveway. Sweepstakes: $5 million at 1 in 200 million is worth 2.5¢ before you price the stamp. A post-office view of $80 million in stamp revenue shows where the house edge really sits.
Focus groups of six. Only 20 of 64 equally likely opinion splits are 3–3 when the public is tied; 44/64 look "decisive." Marketers who treat a 4–2 room as a mandate are reading noise. Wedding seating: choosing 10 of 100 guests for a top table has so many combinations that listing them at one per second would take on the order of 10,000 billion years.
Series length. A 55 percent team still loses a seven-game series about 40 percent of the time. A 2/3 favourite loses about 1 in 5. An employee who is 60/40 against a rival in a best-of-five still loses about one third of such contests. To make the better side almost sure to win you need series lengths like best of 23 or, in extreme imbalances, 269. Short playoffs are entertainment first and measurement second.
German Lotto 6/49 repeated a combination in 1995 that had appeared in 1986; after 3,016 draws some repeat was about 28 percent likely. Canadian "car a day" lotteries produce repeat-winner headlines at rates the combinatorial base rate already expects.
Worked counting and expectation
Birthday approximation. P(no shared birthday) ≈ exp(-n(n-1)/2×365). At n=23 the complement crosses 1/2; at n=50 it exceeds 0.97. The surprise is psychological: people compare to 365, not to C(n,2).
Yankees down 0–2. Remaining best-of-five equivalent paths: win 4 before the opponent wins 2 more... under even games, 6 of 32 elementary paths give a comeback (≈19 percent). Sportstalk that treats a 2–0 hole as destiny overstates; sportstalk that treats it as trivial understates.
Virginia cover. Combinations 7,059,052; pot $27,918,561; bulk ticket economics near $3.95 face-equivalent; syndicate of about 2,500 investors; about 5 million tickets bought; win secured; split-adjusted value near $23.4M. Expectation flipped because the pot outran the cost of covering enough of the space - a rare design failure, not a consumer tip.
Pascal's triangle row 6. Entries 1,6,15,20,15,6,1. A six-person focus group has 20 ways to split 3–3 and 44 ways to look unbalanced. Marketers reading 4–2 as a mandate are reading the row wrong.
Apply counting and expectation
Roadmaps. A "best of five" bake-off between two vendors who differ only slightly will mis-crown the weaker one about a third of the time if quality is 60/40. Prefer longer trials or larger scorecards.
Promotions. Focus groups of six are structural noise machines when the public is split. Require pre-registered decision rules.
Budget bets. Price rare fines, outages and sweepstakes-like upside with expectation, not with the vivid headline payoff.
More from Galileo, de Mere and Pascal's wager
Galileo Galilei answered a grand-duke puzzle about three-dice totals while also working on pendulums (1583 observation tradition). The 27 vs 25 microstate counts for totals 10 and 9 are a textbook of why intuition about "near" outcomes fails. The Chevalier de Mere brought Pascal the points problem and a dice complaint about the frequency of a six in four rolls versus a double-six in twenty-four - another counting gap. Fermat in Toulouse answered with enumeration that matched Pascal's combinatorial approach.
Pascal's Pensées wager is decision theory avant la lettre: weight outcomes by probability and by payoff, including infinite payoffs. Whether one accepts the theology, the form of the argument - expectation under uncertainty - is the same form used for parking tickets and lottery syndicates later in the chapter. Parish gambling and medieval chance games supplied the raw material; the 1654 letters supplied the algebra.
COUNT PATHS / PRICE EV. Ordered outcomes (not just combinations) and expected value (probability times payoff, including rare costs) are the chapter's twin tools. Short contests and small focus groups are weak tests of skill.
For one decision this week that has a rare fine, outage or lottery-like upside, write EV = Σ p × payoff. Compare that number to the vivid headline payoff you would have used. Keep the cheaper of the two stories.
Invert Chapter 4: price by headline payoff only. Then lotteries look wise, rare fines look free, de Mere's dice feel like vibes, and Pascal's wager is mysticism rather than decision theory with infinite payoffs.
- Count ordered outcomes, not just combinations.
- Coincidences are common when there are many pairs or many tries.
- Short contests are weak tests of skill; small groups produce false consensus.
- Price decisions by expected value, including rare costs.
5 The Dueling Laws of Large and Small Numbers
Cardano, Galileo and Pascal assumed the probabilities were known. But how do we know a die is fair, or what "70 percent effective" means for a drug? This chapter is about the link between underlying probabilities and observed results, and the two opposite errors we make about it.
Is anything truly random?
Over lunch at Caltech a visiting Hebrew University statistician, Moshe, insisted truly random numbers do not exist: every real die is slightly biased, and enough throws will reveal it. Mlodinow agrees about dice but points to nature: quantum events are genuinely random, and quantum generators now harvest them.
Older attempts failed in instructive ways. Harlem racketeers around 1920 took their daily lottery number from the last five digits of the US Treasury balance, but such cumulative figures obey Benford's law and are not uniform.
In many real-world data sets the leading digit is 1 about 30% of the time, 2 about 18%, falling to 9 about 5%. Simon Newcomb noticed in 1881 that the early pages of log tables were the most worn; Frank Benford rediscovered it in 1938; Ted Hill proved why in 1995.
Lawrence raised $91 million and spent much of it on homes, 20 personal watercraft, 47 cars (Ferraris, Hummers, a Lamborghini Diablo), a $200,000 samurai sword and a cotton candy machine, laundering through shell companies. Forensic accountant Darrell Dorrell ran over 70,000 cheques and transfers against Benford's law; they failed. Lawrence got 20 years in 2003. (Thirteen years of Bill Clinton's tax returns, for the record, passed.)
In 1947 the Rand Corporation built random digits from electronic noise for Monte Carlo calculations. Tests found biases they never fully removed, but in 1955 they published A Million Random Digits. Philosophers split randomness two ways: the frequency view (Peirce, 1896: judge by the long-run results) and the subjective view (judge by whether the process can be predicted). A perfect die is random by the first and not the second; a real die the reverse.
Joseph Jagger breaks the bank
Even Jagger faced the core question: were the hot numbers real bias or chance clustering? Answering it needed calculus.
Bernoulli's golden theorem
The great comet of 1680 turned Basel theologian Jakob Bernoulli to mathematics (he wrongly but bravely claimed comets obeyed natural law), founding the most gifted family in maths history. Inspired by Huygens's probability primer, he wanted probabilities for cases where they are not handed to us: testimony, plague, imperfect dice. They had to be found by observation.
Calculus (Leibniz 1684 and 1686, Newton's Principia 1687) gave him the tools of sequence, series and limit. Zeno's paradox shows the idea: 1/2 + 1/4 + 1/8 + ... never exceeds 1, so the walk to the door ends. John Kerrich, interned by the Germans in Denmark, tossed a coin 10,000 times: 44 percent heads after 100, 50.67 percent after 10,000.
Choose a tolerance of error (say plus or minus 2%) and a tolerance of uncertainty (say 99.9%). Then some finite number of trials guarantees that the observed proportion falls within that error band with at least that probability. Large samples almost certainly reflect the population.
Each yes/no trial is a Bernoulli trial: coin, vote, cure, purchase. Bernoulli's own formula was hopelessly cautious: to pin Basel's 60 percent mayoral support within 2 points at 99.9 percent "moral certainty" it demanded 25,550 people, about all of Basel. Modern methods need only 370 for plus or minus 5 points at the usual 95 percent, and 1,000 gives a 90 percent chance of being within 2 points. Stephen Stigler suggested Bernoulli "literally quit when he saw the number 25,550." He died in 1705; his venomous brother Johann refused to finish the book; nephew Nikolaus delayed; Ars Conjectandi appeared in 1713.
The law of small numbers
Real life rarely gives thousands of trials, so we make the reverse mistake: we treat small samples as representative. Kahneman and Tversky mockingly called it the law of small numbers.
If 60% back the mayor, a poll of 5 finds exactly 3 supporters only about 1 time in 3, even though that is the single most likely result. 2 times in 3 it gives a wrong picture, and about 1 time in 10 it is unanimous.
Apply it to the Fortune 500. Give every CEO a steady 60 percent chance of a good year. Over five years, about 333 of the 500 will show a record that misstates their ability, and about 1 in 10 will have five straight good or bad years by chance alone. Bernoulli's verdict: "One should not appraise human action on the basis of its results." Judging the person instead of the scoreboard takes courage; nobody stands up in a meeting to say "don't fire her, she was on the wrong end of a Bernoulli series."
The belief that a fixed-probability event is "due" or that luck has "run out." Kerrich's coin did not lean towards tails after 44 heads. Mlodinow admits he watched a sweating cruise passenger who was "due" hit the jackpot seconds later, and confesses the illusion is powerful even for him.
Benford's law is now routine in audit analytics and fraud screening, but researchers warn against using it alone, notably in election forensics, where vote counts often violate it for innocent reasons. Randomness has gone commercial: quantum random number generators ship in phones and cloud services, and NIST's randomness beacons and test suites set the bar. Small-sample error remains the root of the replication crisis: large multi-lab replications since 2015 found many famous psychology effects were inflated by small samples, exactly the bias this chapter names.
Extra detail: Benford, Jagger, Peirce and the Bernoulli family
Is anything truly random? Moshe at Hebrew University presses the question; quantum measurement supplies one modern yes. Practical randomness is harder. The RAND Corporation's 1947 Monte Carlo project and the 1955 Million Random Digits book show how much engineering "chance" needs. Charles Peirce (1896) separated frequency probability from subjective degree of belief; both appear in modern use, and confusing them breeds category errors.
Benford's law. Simon Newcomb (1881) noticed dirtier early pages in log tables; Frank Benford (1938) at GE tabulated leading digits; Theodore Hill proved a version in 1995. Leading digit 1 appears about 30 percent of the time, 2 about 18 percent, 9 about 5 percent. Harlem's Treasury lottery around 1920 already showed the skew. Fraud case study: Kevin Lawrence's roughly $91 million scheme, 47 cars, analyst Dorrell scanning 70,000 numbers over a 20-year window ending 2003. Clinton's tax-return digits, by contrast, passed. Benford is a screen, not a verdict.
Bernoulli's golden theorem. Jakob Bernoulli worked toward it for 20 years, inspired partly by a 1680 comet and by Huygens. Calculus from Leibniz (1684/86) and Newton (1687) supplied limits; Zeno's 1/2+1/4+… and Diogenes' walk to the door are the intuition. An urn with 3,000 white and 2,000 black balls is his teaching image. Bernoulli trials are the yes/no atoms. His bound demanded 25,550 people for Basel's 60 percent mayoral support within 2 points at 99.9 percent "moral certainty"; modern methods need about 1,000 for a 90 percent chance within 2 points, or 370 for ±5 at 95 percent. Stigler joked he quit at 25,550. Jakob died 1705; brother Johann was venomous; nephew Nikolaus delayed; Ars Conjectandi appeared 1713.
Law of small numbers. A poll of 5 when 60 percent agree finds exactly 3 supporters only about 1 time in 3; about 1 in 10 it is unanimous. Among Fortune 500 CEOs with a steady 60 percent chance of a "good year," about 333 will show five-year records that misstate ability, and about 1 in 10 will post five straight good or bad years by chance. Bernoulli: "One should not appraise human action on the basis of its results." The gambler's fallacy on a cruise slot is the same error in leisure clothes: independent trials have no memory.
Narrative spine: from Newcomb to Ars Conjectandi
Walk the chapter as a timeline. 1881: Newcomb notices log-table wear. 1920s: Harlem numbers games display Benford skew in the wild. 1938: Benford tabulates leading digits across domains. 1947–55: RAND industrialises random digits for Monte Carlo. 1873 (flashback): Jagger's clerks turn a biased wheel into a bankable edge. 1680–1713: Jakob Bernoulli pursues the golden theorem through comet year, Huygens, Leibniz, Newton, Zeno and an urn; dies in 1705; Ars Conjectandi appears 1713 via Nikolaus after Johann refuses. Kerrich in a POW camp tosses 10,000 times and watches 44 percent at 100 become 50.67 percent at 10,000 - the law of large numbers in a single scarred notebook.
The pedagogical sequence matters. First admit that true randomness is philosophically contested (Moshe; quantum). Then show that imperfect chance devices still yield law-like frequencies (Jagger; Benford). Then prove that sample size is the bridge from frequency to trust (Bernoulli). Then warn that everyday life offers small samples and that we treat them as large (Kahneman and Tversky's mocking "law of small numbers"). The gambler's fallacy is the small-sample error wearing a leisure suit: the wheel has no memory, and neither does the coin.
| Name | Role |
|---|---|
| Newcomb / Benford / Hill | Leading-digit law |
| Joseph Jagger | Empirical attack on a biased wheel |
| Peirce | Frequency vs belief |
| Jakob Bernoulli | Golden theorem; Bernoulli trials |
| Johann / Nikolaus | Family delay of the book |
| Kerrich | 10,000 POW tosses |
| Kahneman & Tversky | Law of small numbers label |
Apply large and small numbers
OKRs. Five quarterly samples of a noisy KPI will mis-rank managers often. Lengthen the window or shrink the reward attached to a single colour score.
Fraud screens. Benford is a triage tool. Combine it with process knowledge; never convict on leading digits alone.
Casino logic at work. If a metric process is biased (Jagger's wheel), collect the spins. If it is fair, stop paying for streaks.
Kerrich, Jagger and the CEO five-year trap
Kerrich's POW notebook and Jagger's clerks are the same method: record the trials. One proves convergence; the other proves bias. Fortune 500 five-year scoreboards refuse to record enough trials and then fire people for the noise. Bernoulli's line - do not appraise human action on results alone - is the management memo nobody sends. Send it.
LARGE CONVERGES; SMALL LIES. Fix error band and confidence before you call a sample "enough." Independent events have no memory: nothing is "due." Five-year CEO scoreboards mislead about two times in three.
Before you next praise or punish a streak, write (1) the sample size, (2) the error band you would accept, and (3) what "due" would require if trials are independent. If you cannot fill (2), you are in the law of small numbers.
Invert Chapter 5: believe the law of small numbers and the gambler's fallacy. Promote on five-year boards, wait for reversals that are "due," and treat Jagger's biased wheel and Kerrich's fair coin as the same object. Large-number patience becomes optional.
- Large samples converge; small ones mislead, often two times in three.
- Fix your error band and confidence before deciding what sample is "enough."
- Judge ability, not five-year scoreboards.
- Independent events have no memory; nothing is "due."
6 False Positives and Positive Fallacies
The minister who was real
In the 1970s a middle-aged Harvard student explained to his professor that he was the subject of a vast secret experiment run by B. F. Skinner: his wife spoke his thoughts, colleagues foretold his firing. Suing his former employer, he claimed a forgotten eighteenth-century minister's theory showed he was right with probability 999,999 in a million. The company's psychiatrist called the minister a schizophrenic fantasy. The professor deposed that Thomas Bayes (1701 to 1761, minister at Tunbridge Wells, buried in Bunhill Fields) was real, though the student's numbers were not. The sad joke, Mlodinow says, is that the medical and legal experts were the ignorant ones.
P(A given B) is generally not P(B given A). A man who is mentally ill is far likelier than average to think his wife reads his mind, and a man who thinks so is far likelier than average to be ill, but the two numbers differ. Bayes showed how they relate.
Everyday inversions
In a film (Shall We Dance), a lawyer secretly takes up ballroom dancing. His wife reasons that a cheating husband would sneak around, so a sneaking husband must be cheating, confusing P(sneaking given affair) with P(affair given sneaking). A slow-replying boss does not imply your star is falling; she may be busy. Conspiracy theories run on the same swap: a conspiracy would produce these events, so these events prove a conspiracy.
The method is accounting: list the sample space with its probabilities, strike what the evidence rules out, read off what remains.
A family has two children and one is a girl named Florida. Chance both are girls? Not 1/3 but 1/2. Fill a room with 75 million two-child families with at least one girl: 25 million two-girl, 50 million one-girl. If Florida is a 1-in-a-million name, about 50 of the one-girl families stay, and about 50 of the two-girl families stay (each has two tickets). Equal numbers: 1/2. A rare, identifying detail changes the space.
Bayes, Price and Laplace
Bayes never published. His friend Richard Price read "An Essay toward Solving a Problem in the Doctrine of Chances" to the Royal Society on 23 December 1763. Price was no obscure preacher: friend of Franklin, reader of Adam Smith's drafts, and founder of actuarial science, though his life tables underestimated longevity, enriching the Equitable Society and costing the British government on annuities.
Bayes asked Bernoulli's original question the right way round: given the data, what is the underlying probability? His metaphor: roll a ball onto a table, hidden; roll more balls and hear only whether each lands left or right of it; update your estimate of where the first ball lies. Initial estimates are priors, revised ones posteriors.
Posterior odds = prior odds x likelihood ratio. Start with what you knew (base rate), weigh how much more likely the evidence is under one hypothesis than the other, and update. Car insurers do it: a new driver priced as 1/3 high-risk moves towards "low risk" with each claim-free year, which is why good-driver discounts exist.
"999 in 1,000 you'll be dead"
In 1989 Mlodinow and his wife were refused life insurance. His HIV test came back positive and his doctor said the chance he would be dead in a decade was 999 in 1,000, because the test gave false positives only 1 time in 1,000. The doctor had inverted the conditional.
Among low-risk men like him, about 1 in 10,000 of those tested were infected. Per 10,000: one true positive, about ten false positives. So a positive test meant only a 1 in 11 chance of infection; he was in fact healthy. For a group with 1 percent prevalence the same test gives 100 true positives to 10 false, so 10 in 11. The false-positive rate alone means nothing; compare it with the base rate. (A test that calls everyone ill catches 100 percent of tumours.)
Women aged 40 to 50, cancer prevalence 0.8%, false-positive rate 7%, false-negative rate 10%. True chance of cancer after a positive: about 9%. German physicians' median estimate: 70%; a third said 90%. In a US group, 95 of 100 said about 75%.
The prosecutor's fallacy
Mary Decker Slaney, 1983 world champion, was found guilty of a doping offence in 1996 on a test whose false-positive rate may have been 1 percent. That does not mean 99 percent guilt. If 1 in 10 athletes dope and the test catches half of them, then per 1,000: 50 guilty positives and 9 innocent ones, so guilt is 50/59, about 85 percent, roughly the confidence that a die won't show a 1. With 90,000 tests a year, many innocents get condemned.
Two of Clark's babies died, at 11 and 8 weeks, reportedly of SIDS. Paediatrician Sir Roy Meadow testified that the odds of two SIDS deaths were 73 million to 1 (squaring 1 in 8,543), and in 1999 she was jailed. Errors: the deaths are not independent (genetic and environmental links; a BMJ editorial estimated 2.75 million to 1), and the relevant comparison is double SIDS versus double murder, both rare. One mathematician found double SIDS about 9 times likelier. The Royal Statistical Society protested. A withheld finding of bacterial infection in the second child finally freed her after nearly 3.5 years.
| Defence statistic | Relevant statistic |
|---|---|
| Of 4 million battered women a year, only 1,432 (1 in 2,500) are killed by their partner | Of battered women who were murdered, about 90% were killed by their abuser |
| True, convincing, irrelevant | Never mentioned at trial |
Alan Dershowitz later wrote that only witnesses swear to tell the whole truth.
From probability to statistics
Laplace, reinventing Bayes from 1774 onward, turned it into a general method. A political survivor, he examined the young Napoleon, praised the Republic, became a count under the Empire, dedicated his 1812 treatise to "Napoleon the Great" and in 1814 predicted the fall of would-be universal empires. His work marks the line: probability predicts data from known odds; statistics infers the odds from data.
Mlodinow's 1-in-11 story reflects 1989 assays; modern combined HIV tests have far lower false-positive rates and positives are always confirmed with a second test. The mammography lesson stands: the US Preventive Services Task Force (2024) now recommends screening every two years from age 40, and its materials stress that most positive screens are false alarms. Teaching with natural frequencies ("10 in 1,000"), championed by Gigerenzer, is now standard in risk communication. Sally Clark died in 2007; her case led to a review of other convictions involving Meadow's evidence. In 2023 to 2025, UK debate over statistical and circumstantial evidence in the Lucy Letby case revived calls, including from the Royal Statistical Society, for statistical guidance in courts. Bayesian methods also now underpin spam filters, medical AI calibration and A/B testing platforms.
Extra detail: Bayes, Price, Florida, Slaney and Laplace
Thomas Bayes (1701–1761), minister at Tunbridge Wells, buried at Bunhill Fields, left a manuscript that Richard Price read to the Royal Society on 23 December 1763 and published in 1764; Price received £100. Price's actuarial work for Equitable underestimated life expectancy, a reminder that Bayesian tools do not automate wisdom. The two-ball table metaphor (prior positions, posterior after a hit) is Mlodinow's teaching picture of priors updating into posteriors.
Everyday Bayes. A Harvard student sees Skinner and invents a conspiracy; Bayes would not assign 999,999/1,000,000 without a prior. A dance-film date (Shall We Dance) and a boss's ambiguous emails are ordinary settings where people skip the prior and leap. Car insurance: a 1/3 high-risk prior updates with driving history. The Florida girl problem: after a girl is found, the chance a specific missing child is hers moves from a tiny base rate (1935 figure 1/30,000; modern illustrations use 1 in a million) toward 1/2 once you condition on the filtered families still in play among roughly 75 million households with rooms to search. Numbers are illustrative; the structure is the lesson.
Mlodinow's 1989 HIV result. Prevalence about 1 in 10,000 among low-risk tested men; false-positive rate about 1/1,000 → about 1 in 11 chance a positive is a true infection. In a group with 1 percent prevalence the same test yields about 10 in 11. Mammogram illustration: prevalence 0.8 percent, FP 7 percent, FN 10 percent → about 9 percent chance of cancer after a positive; German physicians' median guess 70 percent; US group 95/100 said about 75 percent.
Mary Decker Slaney (1996): FP near 1 percent does not mean 99 percent guilt; under plausible doping base rates guilt might be about 50/59 ≈ 85 percent. With 90,000 tests a year, many innocents are caught in the net. Sally Clark (1999): Meadow's 1 in 8,543 squared to 73 million; BMJ editorial order 2.75 million; RSS protested; double SIDS about 9× likelier than double murder in one analysis; released after ~3.5 years when infection evidence surfaced. O. J. Simpson: Dershowitz's 4 million battered women and 1,432 partner killings (1/2,500) was true and irrelevant beside the fact that about 90 percent of murdered battered women are killed by their abuser.
Laplace reinvented and extended Bayes from 1774, examined young Napoleon, dedicated his 1812 treatise to "Napoleon the Great," and by 1814 was adjusting to new masters. He marks the line: probability predicts data from known odds; statistics infers odds from data.
Narrative spine: from Bayes's notebook to the courtroom
1701–61: Bayes preaches and thinks in Tunbridge Wells. 1763–64: Price rescues the manuscript, reads it, publishes, collects £100, and mis-prices longevity at Equitable. 1774–1814: Laplace turns the idea into a method and a political weather vane. 1989: Mlodinow's own HIV false positive becomes the book's central natural-frequency tree. 1996–99: Slaney and Sally Clark show prosecutor fallacies with careers and liberty on the line. O. J.: Dershowitz's true 1-in-2,500 statistic distracts from the relevant 90 percent.
The chapter's discipline is a single sentence: P(evidence | guilt) is not P(guilt | evidence). Everything else is commentary. Base rates turn accurate tests into mostly false alarms when the condition is rare. Competing explanations (double SIDS vs double murder) beat one explanation waved against zero. Natural frequencies - "of 10,000 people like you…" - are the communication patch that makes the arithmetic feel like common sense again.
Posterior odds = prior odds × likelihood ratio. Skip the prior and you are doing theology with percentages. Skip the likelihood ratio and you are doing rumour with base rates.
More on Price, priors and political Bayes
Richard Price was not only Bayes's editor; he was an actuarial pioneer whose Equitable work underestimated survival. The irony is pedagogical: the person who midwifed inverse probability still needed better data. Mlodinow's dance-film and boss-email vignettes show everyday prior neglect - we jump to a favoured hypothesis because it is available, not because the likelihood ratio demands it. The Florida missing-child illustration (period base rates from 1935 onward) is a worked prior that moves a tiny number toward 1/2 only after aggressive conditioning. Treat the arithmetic as structural, not as a modern missing-persons manual.
NEVER SWAP IF/GIVEN. P(evidence given guilt) is not P(guilt given evidence). Base rates rule; rare conditions turn accurate tests into mostly false alarms. Think in natural frequencies: of 10,000, how many?
Recast one medical, legal or fraud-alert claim as a natural-frequency table for 10,000 people (or cases). Fill true positives, false positives, and the posterior. Speak the posterior aloud before you trust the headline percentage.
Invert Chapter 6: swap the conditionals. Then DNA matches, HIV screens and "999 in 1,000" rhetoric become verdicts, the prosecutor's fallacy is advocacy, and Bayes is a curiosity for actuaries. Prior odds disappear from the room.
- Never swap "if" and "given": P(evidence given guilt) is not P(guilt given evidence).
- Base rates rule: rare conditions turn accurate tests into mostly false alarms.
- Compare competing explanations, not one explanation against zero.
- Think in natural frequencies: of 10,000, how many?
7 Measurement and the Law of Errors
The essay that scored 93
Mlodinow confesses he rewrote his son Alexei's English essay late one night, and it came back with a 93, "more of an A- than an A," as Alexei consoled him. Another professional writer's daughter got a B for his work. Two of Alexei's friends handed in identical essays and received 90 and 79. The point: a grade is not a description of quality but a measurement of it, and every measurement carries random error.
Elections are measurements too. Washington's 2004 governor's race: the Republican led by 261 of about 3 million votes; the recount cut it to 42; a hand count put the Democrat ahead by 10, then by 129 after 700 "lost" ballots turned up. The change between counts (219) was bigger than the margin. A 42-vote edge is 1 vote in 70,000. When results fall inside the error, Mlodinow suggests, accept them or toss a coin rather than recount forever.
From golden numbers to the mean
Most eighteenth-century scientists did not average. They picked one "golden number" and hid the rest, treating variation as moral failure. Newton was an early exception. Precision astronomy and the new French experimental physics of the 1780s (Laplace, Coulomb, and Lavoisier, guillotined in 1794 because "the republic has no need of scientists") forced a theory of measurement, which became mathematical statistics.
Numbers that look precise
- A radar reading of 39 in a 35 zone sticks, though radar guns vary by several mph.
- Students can often gain 30 SAT points just by retaking the test.
- A jobless rate moving from 4.8 to 4.7 percent made headlines; Barron's Gene Epstein noted a 0.1 move is indistinguishable from noise.
- At Clarion University, 120 papers each graded by eight faculty differed by nearly a full grade on average. Trained raters at Iowa State agreed only about half the time on a 4-point scale. The College Board expects only 92 percent of SAT essays to fall within 1 point between two raters.
Robert Parker's 100-point scale (1978) helped build a $20 billion US market, where a 90 instead of an 89 transforms sales. Yet taste blends tongue and nose across 600 to 800 volatile compounds, and trained tasters can identify only 3 or 4 components in a mixture. Expectation bends perception: white wine dyed red tastes "red" to oenology students; in a 2008 fMRI study the same wine labelled $90 beat its $10 label and lit up pleasure circuits more; 21 of 30 cola drinkers "confirmed" their brand with swapped bottles. In triangle tests (pick the odd one of three) experts succeeded only about two-thirds of the time; they disagreed on 9 of 12 components. The 1999 Mitchelton Blackwood Park Riesling was one guide's wine of the year and another's worst vintage in a decade. Critics admit the system is "nonsensical," but numbers sell.
Mean, spread and the error law
Fifteen critics all scoring 90 and fifteen scoring from 80 to 100 share a mean of 90 but say different things. The standard deviation captures the difference: 0 for the first set, about 6 for the second, so the wine is "somewhere between 84 and 96."
Bessel listed eleven sources of telescope error in 1838. Daniel Bernoulli (1777) compared observational errors to an archer's arrows clustering around the bullseye. The deep claim of statistics is that errors follow a universal law, so finding Jupiter, the weight of a loaf and a wine's "true" score are the same problem. (Not always: critics split into red-lovers and white-lovers would give two peaks.)
De Moivre's curve and Laplace's theorem
Abraham De Moivre found the curve in 1733 while approximating the giant numbers deep in Pascal's triangle (one entry in row 200 has 59 digits). Plot the rows as bars and the tops approach a bell. Today this is the normal (Gaussian) distribution, set by a mean (location) and a standard deviation (spread).
If a poll's margin is plus or minus 5 points, then 19 times in 20 a repeat would land within 5 points of the truth, and 1 time in 20 it will not. Roughly, 100 people is too few; 1,000 gives about 3 points. Six real polls of presidential approval in September 2006 read 37, 39, 39, 40, 42, 42: normal scatter. CNN's 2004 "Bush bounce" of 2 points sat inside a 3.5-point margin: "apparently," in CNN-talk, meant "apparently not."
Mlodinow's chart compares 300 students guessing coin flips with 300 mutual fund managers counting years out of 10 that they beat the average: the two bell curves nearly coincide (a point he returns to in chapter 9). In life we usually see one data point: one record season, one hot broker, one hit product. A sample point is just a sample point; read it against the spread that produced it.
Gauss connected the curve to measurement errors in astronomy, though by his own later admission his proof was invalid. Laplace, who had just proved the central limit theorem (1810), saw the link: the sum of many small independent random factors is normally distributed. A thousand-gram loaf with a bit more or less flour, moisture and heat varies along a bell. So does an arrow (Mlodinow, aiming a foam arrow at an apple on his son Nicolai's head, missed both and hit a glass of cabernet). By the 1830s scientists accepted every measurement as such a composite.
Measurement noise has its own literature now: Kahneman, Sibony and Sunstein's Noise (2021) documents large disagreements among judges, underwriters and doctors scoring identical cases. Robert Hodgson's studies of California State Fair wine judges (2008 onward) found few could repeat their own scores on the same wine poured blind. The SAT dropped its essay in 2021. Polling misses in the 2016, 2020 and 2024 US elections showed that the published margin of error covers sampling error only; non-response and turnout modelling add systematic error that does not average away, and aggregators now publish wider "total error" estimates.
Extra detail: essay 93, Washington 2004, wine, polls and the bell curve
The essay that scored 93. Mlodinow's son Alexei received 93; friends' identical essays scored 90 and 79. Writing scores are noisy. Clarion: 120 papers, 8 graders, about 1 grade typical swing. Iowa State graders agree only about half the time. College Board: 92 percent of scores within ±1 on a 6-point scale. Parker (1978) found swings near 100 points on some instruments. Software that "ages" essays to 7 years of schooling still leaves human variance.
Washington governor, 2004. About 3 million votes. Machine count: Republican lead 261. Recount: lead 42. Hand count: Democrat lead 10. With about 700 "lost" ballots: Democrat lead 129. The 219-vote swing between counts exceeded some margins. At that scale a 1 in 70,000 perturbation matters. Elections near ties are measurements with error bars wider than the headline.
Wine and taste. US wine is a roughly $20 billion market. Valentin Bianchi scores 89/90 style spreads show how thin the ranks are. Wines hold 600–800 volatile compounds; experts reliably track about 3–4 components. A 1963 red-dye trick fooled experts into "red" vocabulary for white wine. A 2008 fMRI study of $90 vs $10 wine showed price labels changing brain response. Coke vs Pepsi: 21/30 preferred Pepsi unmarked in one classic test; branding reverses the story. 16 odours: about 1 in 4 misidentified. Triangle tests (1990): about 2/3 correct; experts disagreed on 9 of 12 pairs in one report; matching near 70 percent. A rating SD of 6 around a mean of 90 (scores 80…100) makes one-point ranks theatre.
Mitchelton riesling ranked best and worst in different tastings: the same bottle, different noise draws. Grades, wines and polls share a structure: a signal plus a disturbance, summarised by a mean and a spread.
Narrative spine: from De Moivre to the wine label
1733: De Moivre meets the curve in Doctrine of Chances. 1777: Daniel Bernoulli's archer. 1810: Laplace's central limit theorem. 1838: Bessel classes errors. Gauss becomes the name on the bell. Twentieth-century applications: SAT retest noise, BLS tiny shifts, Clarion and Iowa State grading disagreement, College Board ±1 bands, Parker's 100-point swings, Washington 2004's 261 → 42 → D+10 → 129 on three million votes, six polls reading 37–42, CNN's 2-point bounce inside a 3.5-point MoE.
Wine is the chapter's human comedy. A $20 billion US market ranks bottles with one-point solemnity while experts track a handful of the 600–800 volatile compounds, fail triangle tests at nontrivial rates, describe dyed white wine with red vocabulary, and show fMRI changes when the price tag lies. Coke/Pepsi unmarked vs branded is the soft-drink twin. Grades, polls and tasting notes are the same object: a mean plus a disturbance. Report the disturbance or you are telling fortunes.
| Sample size (simple random) | Approx. MoE (95%) |
|---|---|
| 100 | ~10% |
| 400 | ~5% |
| 1,000 | ~3% |
| 2,500 | ~2% |
| Treat a "change" smaller than the MoE as a weather report, not a verdict. |
Apply the bell curve and MoE
Performance ratings. If the SD of raters is large, one-point differences are theatre. Report intervals or force coarser bins.
Survey readouts. A 2-point "bounce" inside a 3.5-point MoE is not news. Write the MoE in the same sentence as the change.
Vendor bake-offs. Blind triangle tests beat branded demos when the claim is sensory or UX polish.
More on Lavoisier, radar and false precision
Lavoisier's execution in 1794 under the line that the republic had "no need of scientists" sits beside the metric system's birth as a reminder that measurement is political. Radar jokes about "39 in 35" and SAT/BLS tiny shifts show false precision: more digits than the error bar allows. Newton's mean and the "golden number" folklore around grades are cousins. When Alexei's essay earns 93 and identical text earns 79, the instrument, not the writer, moved. Demand the SD of graders before you celebrate a one-point victory.
Grades, wine and polls as the same object
Write every score as x = signal + noise. Alexei's 93 and a friend's 79 on identical text are mostly noise. A wine 90 ± 6 overlaps a wine 88 ± 6 almost completely. A poll at 40% ± 3 overlaps 37% and 42%. The chapter's villains are not graders or sommeliers; they are readers who treat the point estimate as a destiny. De Moivre and Laplace gave the shape of the noise; modern institutions still sell the destiny.
x = SIGNAL + NOISE. Ask for the spread, not just the score. Ignore moves inside the margin of error. Many small random causes add up to a bell curve; a single success or failure is one draw, not the mean.
Pick one grade, wine score, poll or KPI you treated as exact this month. Rewrite it as estimate ± margin (use the book's margin, or a honest guess). For one week, refuse to act on changes that sit inside that band.
Invert Chapter 7: treat point estimates as destiny. Then Alexei's 93 and a friend's 79 on the same essay become character, wine 90 beats wine 88 without overlap, and polls at 40% ± 3 crown a winner. De Moivre and Laplace become irrelevant.
- Every number is a measurement with error: ask for the spread, not just the score.
- Ignore changes inside the margin of error.
- Many small random causes add up to a bell curve (central limit theorem).
- A single success or failure is one draw, not the mean.
8 The Order in Chaos
Jeanne Calment's apartment
In the mid-1960s, Jeanne Calment, about 90, sold her Arles apartment to a 47-year-old lawyer for a monthly payment until her death. She had met van Gogh ("dirty, badly dressed, and disagreeable"). The lawyer should have known that a 90-year-old's conditional life expectancy was about six more years, not "minus ten." She passed 100 and 110 on his money; he died in 1995; she died on 4 August 1997, aged 122. Individual lives are unpredictable; groups are not.
200 million drivers, one creature of habit
About 200 million US drivers drove 2.86 trillion miles one year (14,300 each). The next year, with no plan or quota, they drove 2.88 trillion (14,400 each), and road deaths were 42,815 versus 42,643.
| Year | Event |
|---|---|
| 1086 | Domesday survey counts land and livestock for William the Conqueror |
| 1603 | London's weekly bills of mortality begin in a plague year |
| 1662 | Graunt's Observations: first life table |
| 1835 | Quételet's Sur l'homme: the average man |
| 1869-1888 | Galton: heredity, regression, correlation |
| 1892 | Pearson's chi-square test at Monte Carlo |
| 1905 | Einstein explains Brownian motion |
Graunt and Petty
John Graunt, a wealthy London haberdasher, read the bills of mortality as data. In 1665 only 45 died of starvation (against 68,596 of plague), so Londoners must be feeding beggars. He estimated London at 384,000, not the assumed 2 million; showed growth came from immigration; found population recovered from epidemics within two years; and published the first life table in 1662. He was elected to the Royal Society, then ruined by the Great Fire of 1666 and falsely blamed for it. His friend William Petty, a forerunner of economics, treated subjects as assets (and grabbed Irish land for himself). Graunt's real legacy: a limited sample can reveal a whole population.
Quételet and the average man
Justin Wolfers studied about 70,000 games. The gap between the favourite's winning margin and the betting spread is normally distributed with mean 0 and standard deviation 10.9 points (13.9 in pro football). But among heavy favourites there were too few games won by just over the spread and too many by just under it: the signature of players easing off to "shave points" without losing. In an NCAA poll, 1.5 percent of players knew a teammate paid to play badly.
Quételet hoped for a Newtonian "social physics," with the homme moyen as its prototype. Henry Thomas Buckle's statistical History of Civilization in England spread the idea (Darwin read it; Dostoyevsky read it twice). History judged his maths better than his physics: he found few laws, and not everything is normal. Hit-driven businesses like film (20 percent of movies take 80 percent of revenue) follow fat-tailed distributions where "average" and "standard deviation" lose meaning and mega-outliers arrive every few years.
Galton: regression and correlation
Darwin's cousin Francis Galton measured everything: heads, fidgeting, the attractiveness of women by city, fingerprints (adopted by Scotland Yard in 1901), and the lifespans of the prayed-for (no benefit). In Hereditary Genius (1869) he argued talent was inherited, about 250 per million men eminent, and coined eugenics, later adopted by the Nazis, though there is no evidence he would have approved of their crimes.
Mlodinow rejects Galton's straight line from ability to achievement: persistence matters at least as much as talent (the "ten-year rule"). Effort is in our control, and repeated attempts improve the odds chance offers.
His sweet-pea experiment (1875) showed the offspring of large seeds were smaller than their parents and those of small seeds larger: regression toward the mean, the same in human height. Without it, the tall would grow ever taller. Very tall or brilliant parents should expect less extreme children. He then defined the coefficient of correlation, from -1 to +1, now used everywhere from smoking and cancer to test scores and income. His disciple Karl Pearson invented the chi-square test of fit (1892), showing Monte Carlo roulette was not random and that 26,306 throws of twelve dice fit a probability of 0.3377 rather than 1/3 for a 5 or 6: slightly biased dice, as Moshe predicted.
Brown, Einstein and the drunkard's walk
In 1827 botanist Robert Brown saw pollen granules jiggling endlessly and thought he had found the life force, until he saw the same in asbestos and metals. Maxwell and Boltzmann, inspired by Quételet via Buckle, built statistical physics, but atoms were unseen and the theory made few new predictions. In 1905 Einstein explained Brownian motion: the molecules are too light to move a grain individually, and they hit too often to match the jiggles, but those facts cancel. Visible jumps come only when luck produces a lopsided run of hits from one side, the molecular version of Maris's season. The paper became his most cited, settled the reality of atoms, and made the random walk a basic process of nature, seen in foraging mosquitoes, polymers, quantum particles and stock prices.
Calment remains the oldest verified human; a 2018 claim that her daughter took her identity was examined and rejected by French researchers in 2019. On talent, a 2014 meta-analysis (Macnamara, Hambrick, Oswald) found deliberate practice explains only a modest share of performance differences, and "grit" adds little beyond conscientiousness in later meta-analyses: success is effort, talent and luck, closer to Mlodinow than to either Galton or the "10,000 hours" story. Forensic statistics has grown into routine sports integrity monitoring, where betting-market anomaly alerts flag suspect matches. Gigerenzer estimated about 1,600 extra US road deaths in the year after 9/11.
Extra detail: Graunt, Quetelet, Galton, Pearson and the drunkard's walk
Jeanne Calment struck a reverse-mortgage-style deal in the mid-1960s at about age 90 with a lawyer aged 47; he died in 1995, she on 4 August 1997 at 122. Individual lives are wild; aggregates are tame. Kant (1784) already sensed that history might show order in the mass. About 200 million US drivers: miles 2.86T → 2.88T, fatalities 42,815 vs 42,643 - near-stable totals from chaotic trips averaging about 14,300–14,400 miles.
Graunt and Petty. Bills of mortality begin 1603. John Graunt (1662 life table; 1665 plague year): 68,596 plague deaths, 45 starved - so few that he inferred Londoners fed beggars. He put London at about 384,000, not the assumed 2 million. William Petty pushed political arithmetic. The Great Fire (1666) wiped parish data and forced new counting habits. Domesday is the distant ancestor: rulers count when they must tax and conscript.
Galton. Fingerprints to Scotland Yard 1901; studies of prayer; Hereditary Genius (1869); about 250 eminent per million; the ugly programme of eugenics; a ten-year rule of persistent work; sweet-pea experiments (1875) that revealed regression; the correlation coefficient. Pearson's chi-square and Monte Carlo (1892): 26,306 tosses of 12 dice, p≈.3377. Robert Brown (1827) saw pollen jitter; Einstein (1905) explained Brownian motion - among his most cited papers - as a molecular drunkard's walk. The chapter's title image is that walk: no plan at the step, predictable spread at the scale.
Narrative spine: from bills of mortality to Einstein's pollen
1603: bills of mortality. 1662–65: Graunt's life table and plague counts (68,596 plague, 45 starved). 1666: Great Fire. 1784: Kant's hint of order in the mass. 1796–: Quetelet's average man, constant murders, Scottish chests, short conscripts. 1827: Brown's jittering pollen. 1869–1901: Galton's genius counts, sweet peas, regression, correlation, fingerprints. 1892: Pearson's dice. 1905: Einstein's Brownian motion. Mid-1960s–1997: Calment outlives the lawyer. Modern echoes: Poincaré's baker, options backdating, Wolfers's games, retail shrink, post-9/11 road deaths, film's non-normal 20/80.
The drunkard's walk is the unifying picture. Single steps are unpredictable; the cloud of many steps has a radius that grows like the square root of time. Social statistics is the same claim about people: the individual path is a drunkard; the city's death count is a cloud with a measurable radius. Galton's regression is what happens when you select an extreme step and take another: the next step falls back toward the mean of the cloud.
Apply social statistics
Ops dashboards. City-scale rates (crashes, tickets, churn) can be stable while any one district looks wild. Manage the aggregate and investigate outliers with a Poisson null in mind.
HR. Average commuting injuries or sick days can be forecast even when no individual's year can. Do not moralise every personal spike.
Strategy. Film-style 20/80 payoffs need power-law thinking, not bell-curve planning. Match the distribution to the domain.
From Domesday to Brownian motion in one breath
Rulers count when they must tax (Domesday). Cities count when they must bury (bills of mortality). Graunt turns burials into a life table. Quetelet turns crimes and chests into l'homme moyen. Pearson tosses dice; Einstein explains pollen. The drunkard's walk is the mathematical object that makes the whole sequence one idea: unpredictable steps, predictable clouds.
AGGREGATE STABLE; PERSON NOISY. Plan on the aggregate. Departures from the expected distribution are evidence of manipulation. Not all data are normal: hit businesses have fat tails. Extremes regress; correlation measures how tightly two things move together.
Name one metric you currently plan or judge at the person level (commutes, sick days, one film's opening, one hire's first year). Rewrite the plan at the aggregate level, and note which distribution (bell or fat-tailed) the domain actually has.
Invert Chapter 8: expect individuals to be as predictable as averages. Then HR moralises every spike, strategy uses bell-curve planning for 20/80 film payoffs, and Quételet's "average man" becomes a person you can hire. Graunt's life table collapses into biography.
- Individuals are unpredictable; aggregates are stable: plan on the aggregate.
- Departures from the expected distribution are evidence of manipulation.
- Not all data are normal: hit businesses have fat tails.
- Extremes regress; correlation measures how tightly two things move together.
9 Illusions of Patterns and Patterns of Illusion
Tables that moved themselves
In 1848 the teenage Fox sisters heard rappings, worked out a code with "the spirit," and launched spiritualism. By 1853 table-turning was a craze. Physicians found tables moved only when sitters formed a shared expectation; split the group's expectations and the table stood still. Michael Faraday proved the sitters' hands moved before the table and built an indicator that showed them what they were doing: once they could see it, "the power is gone." The motion began as random fidgeting, sitters saw a pattern in it, and expectation did the rest. Perception, Faraday concluded, is an act of imagination.
Even sight is reconstruction: the raw retinal image has a blind spot and sharp resolution over only about 1 degree, the width of a thumb at arm's length. The brain fills gaps and gives us the illusion of clarity. We fill gaps in other data the same way.
Significance testing and its limits
To protect against false patterns, R. A. Fisher (1920s) built significance testing: compute the chance of seeing data this extreme if the "nothing is going on" hypothesis is true; if it is low (say 3 percent), reject it. But test 100 non-psychics or 100 useless drugs and a few will pass by chance, which is one reason small studies contradict each other. In daily life we use gut instinct instead: one bad stove plus one friend's story, and a brand is blacklisted.
A $12.5 million 2006 NEJM trial found glucosamine plus chondroitin no better than placebo for knee arthritis pain. An eminent doctor closed a radio discussion of it by citing a colleague's cat that "cannot get up in the morning" without the supplement. One vivid anecdote outweighed a trial in his mind.
Streaks are what randomness looks like
Kahneman and Tversky called our shortcuts heuristics and their systematic errors biases. Research concluded that "people have a very poor conception of randomness; they do not recognize it when they see it and they cannot produce it when they try."
George Spencer-Brown calculated that a random string of 101,000,007 binary digits should contain about 10 separate runs of a million zeros. Apple's first iPod shuffle played the same artist back to back, users complained, and Steve Jobs made it "less random to make it feel more random." Random processes produce clumps, runs and streaks; it is perfectly even output that would be suspicious.
Mlodinow prints 200 coin tosses with a run of six tails and opening blocks of four. Such runs are expected, but if they were down days on the market, experts would call the market "jittery."
Wall Street and the hot hand
Stock returns are so close to random that, after costs and without inside information, deviations cannot be exploited. Yet analysts averaged about $3 million a year in the late 1990s. Barron's roundtable superstars matched the market (1995); Wall $treet Week picks lagged it (1987, 1997 studies); a study of 153 newsletters found no significant stock-picking skill. Wharton students, given four S&P 500 index funds, mostly failed to pick the cheapest, though an extra 1 percent a year in fees can cut a retirement pot by a third to a half.
Sports columnist Leonard Koppett's system called the stock market correctly 18 of 19 years (1979 to 1998, missing only 1990): if an original NFL team won the Super Bowl, the market would rise. Nobody credits him with skill. Without the explanation, he could have been hailed as a genius.
Bill Miller's Legg Mason Value Trust beat the S&P 500 for 15 straight calendar years, earning "Fund Manager of the Decade" and quoted odds of 372,529 to 1. Mlodinow shows why the number is wrong:
- The streak was a calendar artefact: in 2004 his fund rose under 12 percent while the average S&P stock rose over 15 (he beat only the cap-weighted index), and he lost in more than 30 twelve-month windows that didn't start on 1 January.
- Odds for one named person over 15 named years are 1 in 32,768. But with 1,000 managers tossing coins from 1991, the chance someone goes 15 for 15 is about 3 percent; across the 40 years and thousands of managers considered, about 3 in 4.
If no one had a Miller-like streak, you could complain that the managers were doing worse than coins. Nobel economist Merton Miller: with 10,000 people picking stocks, "one in 10,000 is going to score, by chance alone."
| :::vs Hot hand: perception versus data (1985 view) | Belief | Tversky, Gilovich and Vallone's finding | |
|---|---|---|---|
| A player who has hit is likelier to hit next | 76ers' shots, Celtics' free throws, Cornell players: success after success no likelier than after a miss | Long streaks mean something | Purcell: except DiMaggio's 56 games, nothing in baseball beyond coin-toss models; Orioles' 0-21 start in 1988 included |
Clusters in space
The same holds for maps. A pixelated photo still reads as a baby's face; rearrange the 200 coin tosses in a 5-by-40 grid and corner clusters appear. In 1944, V-2 impacts on London seemed clustered, prompting theories about spies and precision guidance. R. D. Clarke (1946) split south London into 576 half-kilometre squares: 229 had no hits and 8 had four or five, exactly what a random (Poisson) pattern predicts.
California's Raymond Neutra estimated that across 5,000 census tracts and dozens of cancer types, about 2,750 tracts would show statistically significant but random excesses. Worse, residents usually draw the boundary after noticing cases, like the sharpshooter who paints the target around his bullet holes. Thousands of reports a year yield hundreds of studies and almost no confirmed environmental causes; a Minnesota epidemiologist called them "a complete waste of taxpayer dollars." (Real clusters do exist; the point is that most reported ones are noise.)
Why we see patterns: control and confirmation
We crave control. Bettelheim observed that keeping some independent action helped camp survival; rats stripped of control gave up and died. In Ellen Langer's nursing-home study, residents allowed to arrange their rooms and care for a plant scored higher on well-being, and 18 months later had a 15 percent death rate against 30 percent for those given no control.
But if events are random we are not in control, so the need for control fights the recognition of chance. People pressing dummy buttons believe they steer random lights. Langer's subjects felt more confident against a nervous opponent in a pure-chance card game. Yale students given rigged 50 percent coin-guessing scores, some with early streaks, reported that practice would help (40 percent) and distraction would hurt (25 percent); those with early streaks rated themselves better. In her lottery, people who chose their card asked more than four times as much to sell it.
In business the illusion makes leaders the cause of everything. A Columbia and Harvard study of firms forced to change management after bad spells found no average improvement in operating performance over the next three years. The illusion grows with strategising, active involvement and competition, the stuff of corporate life.
Told that 2, 4, 6 fits a rule, most people test 4, 6, 8 and 20, 24, 30, then announce "increasing even numbers." The rule was any increasing numbers. Few try a sequence that could disprove their guess. Francis Bacon saw it in 1620. In a death-penalty study, supporters and opponents read the same mixed evidence, rated the studies agreeing with them higher, and both came away more convinced. Teachers, interviewers, clinicians and people judging minorities all show the same tilt.
Mlodinow's remedy: accept that chance produces patterns, question your perceptions, and spend as much time looking for evidence you are wrong as for reasons you are right.
The hot hand came back. In 2018 Joshua Miller and Adam Sanjurjo showed that the 1985 method was biased: in finite sequences, the chance of a hit right after a run of hits is expected to fall below the player's average even for a coin. Corrected, the Cornell data show a modest real hot hand, and later NBA and three-point-contest studies agree it exists but is far smaller than fans think. The deeper lesson holds: people still see streaks far more often than the effect justifies.
Bill Miller's streak ended in 2006, and Value Trust fell sharply in 2007 and 2008. S&P's SPIVA scorecard for year-end 2025 reports that 89.9 percent of US large-cap active funds trailed the S&P 500 over 15 years, and money has kept moving into low-fee index funds. Some classic studies here have weakened: Langer's nursing-home mortality result rests on a small sample, and replications of the death-penalty "attitude polarisation" study find biased evaluation more reliably than actual polarisation.
Extra detail: significance, markets, clusters and the illusion of control
Fox sisters (1848) launched modern spiritualism; table-turning craze 1853; Faraday built an indicator that showed sitters were pushing. Vision science: a blind spot, fovea about 1 degree - we stitch a world from scraps and call it seamless. Pattern hunger is hardware.
Fisher and significance (1920s): a 3 percent example; run 100 tests and a few "significant" results appear by chance. Viking stoves and United Airlines anecdotes show how thin data become folklore. NEJM 2006: $12.5 million glucosamine trial no better than placebo, beside a cat anecdote that had sold millions. Spencer-Brown's 101,000,007 digits contain runs of a million zeros that look impossible until you count the trials. Apple made the iPod shuffle "less random to feel more random" after users complained about true random clumps. Reichenbach (1934) and classroom sequences of 200 tosses train the eye to expect clumps.
Markets. Late-1990s analysts averaged about $3 million; Barron's roundtable 1995 matched the market; Wall $treet Week picks 1987/1997 lagged; 153 newsletters showed no reliable skill. Wharton students failed to pick the lowest-fee index funds. A 1 percent fee can cut retirement wealth by about one third to one half. Koppett's Super Bowl indicator went 18/19 (1979–1998, bar 1990) - a parable of data mining. Bill Miller's Legg Mason Value Trust beat the S&P for 15 years; quoted odds 372,529 to 1; Consilient Observer variants 1 in 4,096 / 477,000 / 2.2 billion depending on framing. In 2004 he gained about 12 percent while the average stock gained about 15 percent; he had lost in more than 30 twelve-month periods; calendar artefacts matter. Among 1,000 managers, 1 in 32,768 for a long streak still yields a few stars (~3 percent); stretch to 40 years and ~75 percent chance someone looks legendary among 6,000 managers. Merton Miller's line: if you torture the data long enough, they confess.
Illusion of control. Bettelheim; rats pressing levers; Langer's nursing-home plants: about 30 percent vs 15 percent deaths over 18 months when residents had choices; dummy buttons; nervous rivals; Yale 30 tosses with 25 percent distraction / 40 percent practice effects on confidence; lottery cards people choose sell for about 4× the price of assigned ones. CEO firings studied at Columbia/Harvard: no average improvement over 3 years. Wason's 2-4-6 task; Bacon (1620) on confirmation; death-penalty studies that polarise; teachers, interviewers and counsellors who see stereotypes confirmed. We prefer a story of control to a ledger of chance.
Narrative spine: from Fox sisters to CEO firings
1848–53: Fox sisters and Faraday's table-turning indicator. 1620 / 1920s: Bacon on confirmation; Fisher on significance. 1934: Reichenbach on sequences. 1946: Clarke's V-2 Poisson. 1985: Tversky on the hot hand. Late 1990s–2000s: analyst pay, Barron's, Wall $treet Week, 153 newsletters, Wharton fee blindness, Koppett's Super Bowl toy, Miller's 15-year run and its reframed odds. Langer, Wason, nursing-home plants, lottery-card prices, Columbia/Harvard CEO firing studies: the illusion of control wearing a suit.
The chapter's warning is operational. Before you celebrate a streak, ask how many people were trying and how many ways you could have sliced the calendar. Before you fund a cluster investigation, ask how many tracts you searched. Before you fire a CEO after three bad years, ask what the null of regression plus noise already predicts. Significance without a multiple-comparisons discipline is a machine for minting miracles.
Apply significance and pattern discipline
Experiment platforms. Pre-register metrics; correct for the number of tests; publish the ones that "failed."
Investment committees. Ask how many managers and calendar slices were searched before the star appeared. Reframe Bill Miller-style streaks with the book's alternate denominators.
Health clusters. Count the tracts searched before you paint the bull's-eye. Sharpshooter diagrams belong in the appendix of every cluster memo.
STREAKS ARE NORMAL. Ask how many tried and how the window was chosen before admiring a record. Pay for low fees, not past winners. Feeling in control is not being in control: test ideas by trying to falsify them.
Take one "hot" fund, cancer cluster or winning streak story. Write (1) how many managers, tracts or players were searched, (2) who painted the bull's-eye and when, and (3) one falsifying test you could run. If (1) is unknown, withhold admiration.
Invert Chapter 9: paint the bull's-eye after the shots. Then every cluster is a cause, every streak a genius, significance at 5 percent is a discovery machine, and the illusion of control is a management virtue. V-2 maps and Texas sharpshooters become proof.
- Streaks and clusters are normal outputs of chance.
- Ask how many tried and how the window was chosen before admiring a record.
- Pay for low fees, not past winners.
- Feeling in control is not being in control: test ideas by trying to falsify them.
10 The Drunkard's Walk
Laplace's demon and Lorenz's butterfly
In 1814 Laplace imagined an intelligence that knew every force and position: "nothing would be uncertain, and the future, as the past, would be present to its eyes." Determinism needs three things: known laws, complete data, and enough computing power. Galton and Quételet hoped human affairs met that bar.
In the 1960s meteorologist Edward Lorenz restarted a weather simulation from a printout, typing 0.293 for the stored 0.293416. The forecast diverged wildly. Tiny differences in starting conditions can produce huge differences in outcome: the butterfly effect. His discovery was itself one: a small shortcut opened a field.
Society has no fundamental laws like physics, and people are often irrational. Even with laws, we cannot know the initial conditions precisely. Even with both, the calculations are beyond reach. Max Born: "Chance is a more fundamental conception than causality." Like pollen in Brown's water, we are nudged this way and that, and our responses to the nudges become our path.
Hindsight: why the past looks obvious
Before Pearl Harbor, US intelligence intercepted a request to map the harbour into five areas (decoded 9 October 1941), lost track of Japanese carriers, saw call signs changed twice in a month, and read orders to diplomats to burn codes. On 6 December an army intelligence officer showed one intercept to his boss, who considered it for five minutes and went to dinner.
Mlodinow's answer is a dye molecule in water. It drifts about an inch in three hours. Afterwards we can trace every collision that moved it; beforehand, predicting the path would require tracking every water molecule that might have hit it. Random processes are hard to predict and easy to explain afterwards. Weather and chess work the same way. And the pre-attack signals had innocent readings: similar mapping requests went to agents in Panama, Vancouver, San Francisco and Portland; radio silence often meant ships in home waters; many signs pointed to the Philippines, Thailand or Guam. After the fact, seven congressional committees hunted for culprits. Historian Roberta Wohlstetter: "After the event, a signal is always crystal clear... before the event it is obscure and pregnant with conflicting meanings."
Mlodinow ranks 800 fund managers by 1991 to 1995 returns: a neat descending curve that any analyst could explain. Keep the same order and plot 1996 to 2000: random scatter. Merrill Lynch's Stan O'Neal was a risk-taking genius at $95 a share in spring 2007 and a reckless cowboy, fired, by autumn. A publisher's polished one-, three- and five-year software plans met price wars in year two and a collapsed market in year three.
We can't stop telling stories about the past, but we can treat explanations and prophecies with scepticism, invest in flexibility, confidence, courage and perseverance, and weight direct impressions of people above their trumpeted records.
Normal accidents and lucky breaks
At Three Mile Island (March 1979) about a cup of water leaked from a filter seal into an instrument air line, tripping valves that cut cooling water. Backup pumps ran against valves left closed after maintenance two days earlier; a relief valve stuck open (it had failed at eleven other plants) and a control-room gauge failed to show it. Each fault was routine; together, a partial meltdown. Sociologist Charles Perrow's normal accident theory: in complex, tightly coupled systems, minor faults will sometimes combine by chance into disasters without villains. Flip it around and the same logic explains success: keep trying in a complex world and chance will sometimes line up for you. Economist W. Brian Arthur showed small chance events, amplified by positive feedback, can let a firm with no special edge dominate.
14,341 participants rated and downloaded 48 unknown songs. Eight separate "worlds" could see their own world's download counts; a ninth group saw none, defining each song's intrinsic appeal. Popularity varied wildly across worlds. 52metro's "Lockdown," 26th in intrinsic quality, was No. 1 in one world and No. 40 in another. Early random downloads snowballed.
| Bruce Willis | Bill Gates | |
|---|---|---|
| Before | Seven years acting off-Broadway, working as a bartender | Small software firm |
| Chance event | 1984 LA trip (Olympics or a girlfriend); agent suggests auditions | 1980: IBM's talks with Gary Kildall falter (Kildall's wife refused an NDA); IBM's Jack Sams says "by all means, you get it" |
| Contested moment | Network wanted "Mr. X"; producer Glenn Caron pushed for Willis | Gates buys an OS for about $50,000, licenses DOS to IBM, keeps the rights |
| Early verdict | Moonlighting debuts (March 1985) to low ratings; hit in season 2 | DOS no better than rivals; network effects: "people bought DOS because people were buying DOS" |
Asked about interactive TV on late-night television, Gates gave an ordinary answer. Is he godlike because he earns $100 a second, or does he earn it because he is godlike? The same culture made a TV judge of business talent out of a man whose casino company turned $10,000 (1994) into $636 thirteen years later.
Success makes us see talent
Melvin Lerner found people need to believe the world is just. Observers listened to scripted, equally good workers "Tom" and "Bill," knowing pay went to one at random; whoever got paid was judged the bigger contributor (90 percent for Tom, 70 percent for Bill). Randomly higher-paid partners resisted input more. Watching a "victim" shocked for errors, observers denigrated her more the more she suffered, unless told she was well paid for it. We are biased against those who end up at the bottom.
In David Rosenhan's study, eight sane "pseudopatients" (psychologists, a paediatrician, a painter, a housewife) claimed to hear voices; all were admitted, most diagnosed schizophrenic, and none detected by staff, who logged diary-keeping as "writing behavior." Average stay: 19 days. Teachers give higher marks to identical work from "excellent" students. Odourless, tasteless vodka sells on image; in a New York Times blind tasting of 21 brands, cheap Smirnoff came first.
The Sunday Times sent opening chapters of two Booker Prize winners, V. S. Naipaul's In a Free State and Stanley Middleton's Holiday, to 20 publishers and agents as if from unknowns. All but one rejected them. Stephen King's Richard Bachman novels sold poorly until his identity leaked. Caltech physicist John Schwarz kept string theory alive through a decade of ridicule before a breakthrough made it the hottest field in physics; his equally persistent Berkeley adviser, an S-matrix theorist, never got his break and ended up considered a crank.
Plan with open eyes
Mlodinow's conclusion: ability does not guarantee achievement, and achievement is not proportional to ability. Judge people by their qualities and decisions by the full range of outcomes they could have produced, not the one that happened. And since chance plays a part, one factor is ours to control: the number of at bats. "Even a coin weighted toward failure will sometimes land on success." IBM's Thomas Watson: "If you want to succeed, double your failure rate."
He ends with his mother's sister Sabina, 17, who married a Jewish camp policeman in name only to protect her younger sister, and was gassed with him when the Nazis turned on the police. His mother concluded that planning is pointless. He disagrees: plan, but with eyes open, and appreciate good luck, accept bad luck, and be grateful for the absence of disasters that have not yet come.
Two of this chapter's famous studies need flags. Journalist Susannah Cahalan's The Great Pretender (2019) found serious discrepancies in Rosenhan's records, and many researchers now doubt the study's data; treat it as illustration, not evidence. The MusicLab result (Salganik, Dodds and Watts, 2006) has held up and been extended.
Chaos still caps forecasting near two weeks, but machine-learning weather models (Google DeepMind's GraphCast, 2023; ECMWF's AIFS, operational in 2025) now match or beat physics models at medium range, which supports Lorenz: the gains come from better data and computation, not from escaping sensitivity to initial conditions. Normal accidents keep happening: the July 2024 CrowdStrike update crashed about 8.5 million Windows machines through one faulty content file. Agent-based "talent versus luck" simulations (Pluchino and colleagues, 2018) reproduce Mlodinow's claim that the richest are usually the luckiest of the moderately talented.
Extra detail: Pearl Harbor, TMI, MusicLab, Gates, Willis and just-world bias
Laplace's demon (1814) needed three things: laws, data, computation. Lorenz (1960s) typed 0.293 instead of 0.293416 and met the butterfly effect. Born held that chance was more fundamental than causality. A dye molecule wanders about 1 inch in 3 hours; past/future asymmetry, weather and chess share sensitivity to initial conditions.
Pearl Harbor. Signals piled up: 9 October warnings across five areas; carrier radio silence; call signs changed twice in a month; code burning; a cook's call; Bicknell on 6 December dismissed in five minutes by his boss. Red herrings (Panama, Vancouver…) cluttered the board. Seven congressional committees later reconstructed clarity that nobody had live. Marshall's May 1941 memo called Oahu a "strongest fortress." Tawney and Wohlstetter wrote the lasting lines on hindsight. The lesson is not omniscience; it is reacting under uncertainty rather than demanding prediction that chaos forbids.
Normal accidents. Three Mile Island, March 1979: a cup of water in a polisher, valves left closed 2 days earlier after maintenance, a PORV that had failed in 11 other plants, a gauge that lied - Charles Perrow's "normal accident" in a complex, tightly coupled system. No cartoon villain required.
Positive feedbacks. Brian Arthur: small early leads lock in. MusicLab: 14,341 participants, 48 songs, 8 worlds plus an independent world. 52metro's "Lockdown" ranked 26th on quality without social signal, yet finished #1 in one social world and #40 in another. Same song, different early download noise.
Careers as walks. Bruce Willis, bartender in 1984 LA, Moonlighting; the network wanted another Mr X; Glenn Caron cast him; debut March 1985 with low early ratings before the climb. Bill Gates: IBM August 1980, Gary Kildall's wife and an NDA snag, Jack Sams, $50,000 for QDOS → DOS, royalty structure that kept rights; later wealth on the order of $100/sec across years as richest. Trump casinos: a $10,000 path that becomes $636 in expectation stories - variance and ruin matter more than the brochure. Stephen King as Bachman; John Schwarz and string theory; Chew's S-matrix path; Edison's line on persistence; Watson on doubling your failure rate. Sabina, Mlodinow's aunt, gassed - the book closes the family circle opened in chapter 1.
Just-world and labels. Lerner: Tom/Bill paid 90 percent/70 percent; random pay made observers resist partner input; shock victims were denigrated unless compensated. Rosenhan: 8 pseudopatients, average 19 days to release, "writing behavior" charted as pathology. Teachers grade the "excellent" file higher for the same work. Vodka: NYT 21 brands, Smirnoff tops on brand, not blind chemistry. Booker: Naipaul In a Free State, Middleton Holiday, 20 publishers, all but one rejecting paths that later looked inevitable. 800 fund managers 1991–95 vs 1996–2000: rankings reshuffle like noise. O'Neal at Merrill: genius at $95 in spring 2007, fired by fall. Publishers still demand 1/3/5-year plans that pretend the walk is a highway.
Narrative spine: from Laplace's demon to Sabina
1814: Laplace's demon. 1960s: Lorenz's rounded initial condition. Quantum and Born: chance may be fundamental. Pearl Harbor's unused signals and Wohlstetter's hindsight. TMI as Perrow's normal accident. Arthur's positive feedbacks; MusicLab's 14,341 participants and Lockdown's #1 vs #40. Willis at the bar; Gates and the QDOS fork; King/Bachman; Schwarz; Edison; Watson. Lerner’s just-world lab; Rosenhan's pseudopatients; vodka brands; Booker rejections; fund-manager reshuffles; O'Neal's 2007 arc. Close on Sabina and the Buchenwald bread: the book ends where it began, with a family that knows chance is not a metaphor.
Practical closing rule. In a noisy domain, increase attempts, diversify paths, judge decision quality by information available at the time, and resist the hindsight story that makes every outcome look earned or deserved. The drunkard's walk is not a counsel of despair. It is a counsel of humility plus volume.
Apply the drunkard's walk
Careers. Increase shots on goal; keep rights where you can (Gates/QDOS structure); do not read a quiet year as a character flaw.
Incident response. Train for Perrow-style combinations: closed valves plus a lying gauge plus a routine leak. Hunt couplings, not villains.
Product markets. Early download noise can lock rankings (MusicLab). Design for exploration if you care about quality discovery, not only virality.
ABILITY + CHANCE + ATTEMPTS. Judge decisions by the range of outcomes they could have produced. Distrust neat hindsight (Pearl Harbor). In complex systems, small faults combine (Three Mile Island). Popularity snowballs from chance early leads (MusicLab). Take more at bats.
List three decisions from the last year. For each, write the ex ante range of possible outcomes (not the result you got). Mark which ones you have been judging by the scoreboard alone. Add one concrete way to increase attempts in the noisiest domain.
Invert Chapter 10: judge only by results and hindsight stories. Then Pearl Harbor was "obvious," Three Mile Island needs a villain not a coupling diagram, MusicLab is pure quality discovery, and Gates-style rights structure is destiny rather than path-plus-options. The drunkard's walk metaphor dies.
- The future is hard to predict and the past easy to explain; distrust neat hindsight.
- In complex systems, small faults combine: design for resilience, not blame.
- Popularity snowballs from chance early leads.
- Judge decisions by their range of possible outcomes, and increase your at bats.
Synthesis
25 takeaways
1. Extraordinary outcomes need no extraordinary causes. Maris's 61, Lansing's slump and Miller's streak all fit coin models. 2. Expect regression to the mean before crediting praise, punishment, new managers or turnarounds. 3. Every "and" lowers probability, while making a story more convincing. 4. Multiply only independent probabilities; add only exclusive ones. 5. With two sources of error, the larger one dominates (lab error, not random-match odds). 6. Define the sample space from the real rules, including who knows what (Monty Hall). 7. Conditioning prunes the space: ask exactly what the new information rules out. 8. Count ordered outcomes; coincidences are common when pairs or tries are many (birthday problem). 9. Short contests are weak tests: a 55-45 better team loses a best-of-seven about 4 times in 10. 10. Small groups produce false consensus (two thirds of six-person focus groups mislead). 11. Price choices by expected value, including rare costs and rare windfalls. 12. Large samples converge; small samples mislead, often two times in three. 13. Nothing is "due": independent events have no memory. 14. P(A given B) is not P(B given A). Swapping them is the prosecutor's fallacy. 15. Base rates decide what a positive test means. 16. Compare competing explanations, not one explanation against nothing (Sally Clark). 17. Every measurement has a spread; ignore moves inside the margin of error. 18. Many small causes add to a bell curve; but hit businesses have fat tails. 19. Aggregates are stable even when individuals are not: plan on the aggregate. 20. Departures from the expected distribution can expose fraud (Benford, Quételet, Poincaré, Wolfers). 21. Random processes make streaks and clusters; even output would be suspicious. 22. Ask how many tried and how the window was chosen before admiring a record. 23. Feeling in control is not being in control; test beliefs by trying to falsify them. 24. The past is easy to explain and the future hard to predict: distrust hindsight and long-range plans. 25. Judge people and decisions by the range of outcomes possible, and take more at bats.
Frameworks at a glance
| Framework | Origin | Question it answers | Use it when |
|---|---|---|---|
| Sample space | Cardano | What are all the equally likely outcomes? | Any confusing odds problem |
| Counting / Pascal's triangle | Galileo, Pascal | In how many ways can it happen? | Series, quizzes, focus groups |
| Expected value | Pascal | What is the average payoff? | Pricing, bets, risk trade-offs |
| Law of large numbers | Bernoulli | How many trials until results reflect odds? | Sample sizing, A/B tests |
| Law of small numbers (error) | Kahneman, Tversky | Am I trusting too little data? | Reviews, KPIs, early results |
| Bayes' theorem | Bayes, Laplace | Given evidence, how should belief change? | Diagnosis, fraud alerts, forensics |
| Normal distribution, CLT | De Moivre, Gauss, Laplace | How much spread is noise? | Measurement, polls, quality control |
| Regression and correlation | Galton, Pearson | Will extremes persist? How linked are two things? | Performance, forecasting |
| Chi-square / significance | Pearson, Fisher | Does data fit a model or chance? | Experiments, audits |
| Random walk | Einstein | How does unplanned motion spread? | Prices, diffusion, careers |
| Normal accidents | Perrow | Why do complex systems fail without villains? | Operations, resilience design |
| Cumulative advantage | Arthur, Salganik | How do small early leads snowball? | Platforms, markets, culture |
A closer look
Strong: the mathematics (sample spaces, Bayes, large numbers, the bell curve, regression) and the legal and medical cases (Collins, Clark, mammography) are textbook-solid. The fund-manager argument has only grown stronger; SPIVA's year-end 2025 data show about 90 percent of large-cap active funds trailing the index over 15 years.
Revised: the hot hand is not a pure fallacy; Miller and Sanjurjo (2018) found a small real effect hidden by a statistical bias. Rosenhan's pseudopatient study is now doubted. Langer's nursing-home mortality numbers come from small samples. The death-penalty "polarisation" effect replicates weakly.
Underplayed: Mlodinow occasionally overstates luck. Sherry Lansing and Bill Gates also made choices that mattered, and skill and luck interact: skill decides how many good chances you get and how well you convert them. Michael Mauboussin's "skill-luck continuum" (2012) is a useful complement: chess sits near pure skill, roulette near pure luck, and most careers somewhere in between. The book also says little about fat-tailed risk beyond a paragraph on film, which Taleb's work and the 2008 crisis (which arrived just after publication) made central.
Style: anecdotes are vivid and well chosen; some biographical detail (Cardano, the Bernoullis) relies on colourful but contested sources.
Who's who
| Person | Role in the book |
|---|---|
| Gerolamo Cardano | Gambler-physician; first book on chance; sample space |
| Galileo Galilei | Dice for the grand duke; counting ordered outcomes |
| Blaise Pascal | Problem of points; triangle; expected value; the wager |
| Pierre de Fermat | Pascal's correspondent on the problem of points |
| Jakob Bernoulli | Golden theorem (law of large numbers); Ars Conjectandi |
| Abraham De Moivre | Discovered the bell curve (1733) |
| Thomas Bayes, Richard Price | Conditional probability; Price published it in 1763 |
| Pierre-Simon Laplace | Developed Bayes; central limit theorem; determinism |
| Carl Friedrich Gauss | Linked the normal curve to measurement error |
| John Graunt, William Petty | Founders of statistics; first life table |
| Adolphe Quételet | Social statistics; the average man |
| Francis Galton, Karl Pearson | Regression, correlation; chi-square |
| R. A. Fisher | Significance testing |
| Albert Einstein | Brownian motion as a random walk (1905) |
| Daniel Kahneman, Amos Tversky | Heuristics and biases; regression insight; Linda |
| Ellen Langer, Melvin Lerner | Illusion of control; just-world belief |
| Charles Perrow, W. Brian Arthur | Normal accidents; increasing returns |
| Sherry Lansing, Mark Canton | Studio heads judged on luck-driven results |
| Roger Maris, Bill Miller, Leonard Koppett | Streaks explained by chance |
| Sally Clark, the Collinses, O. J. Simpson | Probability errors in court |
Extra synthesis: a pocket checklist
- Maximising beats matching when one option is clearly more frequent.
- Regression toward the mean follows extremes that mix skill and luck.
- Conjunction never raises probability; good stories often lower it.
- Multiply only for independent events; add only for exclusive ones.
- List the sample space before you compute (Monty Hall, twins, tires).
- Count paths with binomial coefficients; short series are noisy.
- Expectation = Σ probability × payoff (parking, pots, lotteries).
- Law of large numbers: fix error band and confidence, then size the sample.
- Law of small numbers: five-year CEO scoreboards mislead about two times in three.
- Bayes: never swap P(data|hypothesis) for P(hypothesis|data); use base rates.
- Natural frequencies (of 10,000…) cut false-positive confusion.
- Margin of error and the 68-95-99.7 rule discipline polls, grades and wine scores.
- Social aggregates can be stable while individuals stay unpredictable.
- Significance at 5 percent means 1 in 20 nulls cry wolf; many tests imply many wolves.
- Clusters (V-2, cancer) often match Poisson; paint the bull's-eye last and you will always find a sharpshooter.
- Illusion of control sells lottery tickets and CEO firings that do not improve averages.
- Hindsight invents the Pearl Harbor that "should have been obvious."
- Normal accidents arise from routine faults in complex, coupled systems.
- Positive feedback (MusicLab) makes quality and success diverge.
- Judge processes and attempts, not single scoreboard readings; double your trials when the domain is noisy.
A closer look
Mlodinow writes as a physicist-storyteller: examples are chosen for clarity, and some biographical colour (Cardano, Hollywood) is contested in detail by specialists. Hot-hand research since 2018 (Miller and Sanjurjo; 2024 follow-ups) shows the book's "streaks are illusion" line needs nuance in controlled shooting, while fund-manager and hiring streaks mostly do not. HIV and mammography numbers are period-specific; the base-rate logic is not. Read the maths as durable and the newspaper examples as dated illustrations unless an Oct 2026 box says otherwise.
OUTCOMES = A + C + N. Ability, Chance, Number of attempts. Measure with error bars, update like Bayes, distrust streaks and hindsight, design for resilience, take more shots.
From memory, recite the pocket checklist's twenty lines (or as many as you can). Mark the three you fail most often at work. For one week, run only those three as a daily card: regression after extremes, natural frequencies for any test claim, and process-not-scoreboard for one noisy decision.
Invert the whole book: skill explains everything that matters. Then fire coaches after dips, chase five-year fund stars, trust Hollywood anointments, swap courtroom conditionals, treat grades and wine points as destiny, and skip sample spaces. If that inversion still feels natural after the synthesis, reread Chapters 1, 6 and 10 before you close the book.
Outcomes = ability + chance + attempts. Measure with error bars, update like Bayes, distrust streaks and hindsight, design systems that survive bad luck, and keep taking shots.
Glossary
Terms below are the load-bearing vocabulary of the book.
Availability bias. Judging frequency by how easily examples come to mind.
Base rate. How common something is in the population before any test or evidence.
Bayes' theorem. The rule for updating a probability given new evidence: posterior odds equal prior odds times the likelihood ratio.
Bell curve (normal distribution). Symmetric distribution set by a mean and standard deviation; 68, 95 and 99.7 percent of data within 1, 2 and 3 standard deviations.
Benford's law. In many natural data sets the leading digit is 1 about 30 percent of the time, falling to about 5 percent for 9.
Bernoulli trial. A single random trial with two outcomes ("success" or "failure").
Birthday problem. With 23 people, a shared birthday is more likely than not.
Butterfly effect. Sensitive dependence on initial conditions in chaotic systems (Lorenz).
Central limit theorem. The sum of many independent random factors is approximately normally distributed.
Chi-square test. Pearson's test of whether observed counts fit an expected distribution.
Conditional probability. The probability of A given that B has occurred.
Confirmation bias. Seeking and favouring evidence that supports what we already believe.
Conjunction fallacy. Judging "A and B" more likely than "A" alone (Linda).
Correlation coefficient. A number from -1 to +1 measuring how closely two variables move together linearly.
Determinism. The view that the present state fixes the future exactly (Laplace).
Expected value (expectation). The sum of each outcome's probability times its payoff.
False positive / false negative. A test result wrongly saying yes / wrongly saying no.
Gambler's fallacy. Believing an independent event is "due" after a run of the opposite.
Hot-hand fallacy. Seeing skill-driven streaks in random sequences; now known to coexist with a small real hot hand.
Illusion of control. Believing we influence outcomes that are actually random.
Independence. Two events are independent if one occurring does not change the other's probability.
Law of large numbers (golden theorem). With enough trials, observed frequencies almost certainly come close to the true probabilities.
Law of small numbers. The mistaken belief that small samples mirror the population.
Margin of error. The range within which a poll result falls about 95 percent of the time from sampling alone.
Monte Carlo method. Using random numbers to approximate solutions.
Normal accident. Perrow's term for failures that arise from routine faults combining in complex systems.
Prior / posterior probability. Belief before and after seeing evidence.
Problem of points. How to split stakes in an interrupted game; solved by Pascal and Fermat.
Prosecutor's fallacy. Confusing the probability of evidence given innocence with the probability of innocence given evidence.
Random walk (drunkard's walk). A path made of steps in random directions.
Regression toward the mean. Extreme results tend to be followed by less extreme ones.
Sample space. The set of all possible outcomes of a random process.
Sharpshooter effect. Drawing the target after the shots: defining a cluster after seeing the data.
Significance level. The probability of getting results this extreme if the null hypothesis is true; conventionally 5 percent.
Standard deviation. A measure of how widely data spread around the mean.
Extended quick reference
| Term | One-line gloss |
|---|---|
| Availability | Judging frequency by ease of recall |
| Base rate | Prior frequency before new evidence |
| Bernoulli trial | Yes/no trial with fixed p |
| Central limit theorem | Averages of many small effects tend toward normal |
| Conjunction rule | P(A and B) ≤ P(A) |
| Expectation | Probability-weighted average payoff |
| Gambler's fallacy | Belief that independent trials self-correct |
| Hot hand | Belief that success raises P(next success) |
| Law of large numbers | Sample frequencies converge to p |
| Law of small numbers | Treating small samples as representative |
| Margin of error | Half-width of a confidence interval (loosely) |
| Monty Hall | Host-knowledge switching puzzle; switch wins 2/3 |
| Normal accident | Failure from routine faults in complex coupled systems |
| Prosecutor's fallacy | Swapping P(data|innocent) for P(innocent|data) |
| Regression to the mean | Extremes tend to be followed by less extreme trials |
| Sample space | Set of elementary outcomes |
| Significance (p) | P(data as extreme | null); not P(null|data) |
| Drunkard's / random walk | Path of cumulative random steps |
About this book
The Drunkard's Walk: How Randomness Rules Our Lives by Leonard Mlodinow, 2008. This is a chapter-by-chapter study of the book, with summaries, maps, tables and diagrams; short quotations keep their page references. It is shared free for personal study, is not for sale, and is not endorsed by the author or publisher. Please read and buy the original.
Insight Editions · rajeshrmahapatra.com/insight-editions