{
  "id": "probability",
  "title": "Probability",
  "lang": "en",
  "source_lang": "en",
  "summary": "Probability is the mathematical study of uncertainty, assigning numbers between 0 and 1 to events and governing how those numbers combine and update.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "probability theory",
    "chance (informal)"
  ],
  "generated_at": "2026-10-09",
  "model": "pro/claude-haiku-5-5",
  "url": "https://aiwiki.page/en/probability/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/probability/",
      "json": "https://aiwiki.page/zh/probability.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Probability is the branch of mathematics that quantifies uncertainty. It assigns each event a number between 0 and 1, where 0 means the event cannot occur and 1 means it certainly occurs, and it supplies rules for combining, conditioning, and updating these numbers as information arrives. The same formal machinery describes coin tosses, the decay of a radioactive nucleus, the sampling error of an opinion poll, the transmission of a message over a noisy channel, and the predictions of a large language model. Because it is the foundation of statistics, information theory, and much of statistical mechanics and quantum mechanics, probability is one of the most widely applied areas of modern mathematics.\n\nThe subject grew out of problems about games of chance in 17th-century France, most famously the exchange between Pierre de Fermat and Blaise Pascal, and it was developed further by Christiaan Huygens, Jakob Bernoulli, Abraham de Moivre, and Pierre-Simon Laplace. Its modern rigorous form was given by Andrey Kolmogorov in 1933, who defined probability as a measure on a sample space within a σ-algebra of events, using the tools of measure theory. Competing interpretations of what probabilities mean (classical, frequentist, propensity, and subjective or Bayesian) remain philosophically active even though the mathematics is settled.\n\nThis entry first treats the axiomatic foundations and the elementary rules, then random variables, distributions, expectation, and limit theorems, followed by stochastic processes. It next discusses the interpretations of probability and its role in statistical inference, physics, information theory, machine learning, and applied fields such as medicine, genetics, economics, and finance. It closes with classical paradoxes and open questions about foundations and application."
    },
    {
      "heading": "Basic Concept and Everyday Meaning",
      "text": "In everyday speech, probability expresses how confident one can be about something that has not yet happened or has not yet been observed. Tomorrow's weather, the result of a medical test, and the next card drawn from a deck all belong to this category. Probability assigns such an uncertain outcome a number between two extremes: an impossible event has probability 0, and an event that is certain to occur has probability 1. Every value in between expresses partial uncertainty, so the concept lies between the two situations people most often contrast with it, a claim that cannot happen and one that must happen.\n\nThe basic vocabulary is simple. An experiment is any procedure whose result is not fixed in advance, such as tossing a coin or rolling a die. Each possible result of one run is an outcome, and the collection of all outcomes forms the sample space. An event is a set of outcomes, meaning a question about the result that can be answered yes or no after the experiment. For a fair six-sided die, the sample space is {1, 2, 3, 4, 5, 6}, and the event \"the result is even\" is {2, 4, 6}. If each face is equally likely, this event has probability \\(P(\\text{even}) = \\frac{3}{6} = \\frac{1}{2}\\).\n\nThe classical case, in which all outcomes are equally likely, gives the formula \\(P(A) = \\frac{\\text{number of outcomes in } A}{\\text{total number of outcomes}}\\). It works well for dice, cards, and fair lotteries, where physical symmetry justifies the assumption. It cannot, however, cover most real situations. The chance of rain tomorrow, of a patient's recovery, or of an earthquake in a given decade has no natural list of equally likely outcomes and no obvious symmetry to appeal to. Such probabilities must instead come from long-run data, from models, or from a reasoned degree of belief, and these sources are treated in later sections on interpretation and inference.\n\nProbabilities are numbers from 0 to 1, but they are often reported in other forms. A percentage is the probability multiplied by 100. Odds compare the chance that an event occurs with the chance that it does not. If the probability of an event is \\(p\\), the odds in favor are \\(\\frac{p}{1-p}\\). A probability of 0.8 therefore corresponds to odds of 4 to 1 in favor, and a probability of 0.5 to even odds. Odds and probabilities carry the same information but are not identical: odds range from 0 to infinity rather than from 0 to 1, and betting odds quoted by bookmakers usually include a margin, so they do not equal the implied probabilities exactly.\n\nA common misconception is that past outcomes change the odds for independent future events. If a fair coin has landed heads five times in a row, the probability of heads on the sixth toss is still \\(\\frac{1}{2}\\), because the tosses are independent. The streak is unusual as a sequence, since six heads in a row has probability \\(\\frac{1}{64}\\), but that calculation is made before the first toss, and it does not carry over to the next one. The belief that a result is \"due\" after a run of the opposite result is known as the gambler's fallacy. Its reverse, the belief that a streak makes continuation more likely, is sometimes called the hot-hand belief. Both mistakes treat a chance process as if it had memory, which independent trials do not have."
    },
    {
      "heading": "Early History: Games, Gambling, and Reasoning under Uncertainty",
      "text": "Before chance became an object of mathematics, reasoning about uncertainty was mainly qualitative. In ancient Greece, Aristotle distinguished what is necessary from what is merely probable (eikos), and the vocabulary of luck and fortune appeared in discussions of causation and human affairs. These discussions weighed testimony, signs and plausibility, but they did not supply numerical rules for combining chances. The earliest quantitative work on chance came from games. Galileo, between about 1613 and 1623, analysed throws of three dice and observed that some totals occur more often because more combinations of faces produce them. Earlier, Luca Pacioli in 1494 had posed the problem of how to divide the stakes of an abandoned game, and Gerolamo Cardano's treatment of gambling, written around 1564, contained the same kind of reasoning about fair odds, though his solutions were not always correct.\n\nThe decisive step came in 1654. The problem of points, in which a game of chance is interrupted and the pot must be divided fairly, had been raised in Paris by the Chevalier de Méré, and Pascal discussed it with Fermat through letters exchanged in July–September 1654. Their key insight was that a fair division should depend not on the history of the interrupted game but on the ways it could have continued. Fermat enumerated the equally likely continuations; Pascal found a shorter recursive method, later linked to the arithmetic triangle, which he wrote up in a treatise that was not published until 1665. The solutions rested on the concept of expectation, the weighted average of possible gains, which became the central notion of the new calculus.\n\nChristiaan Huygens learned of the problem during a visit to Paris in late 1655, worked out his own solution, and published the first printed treatise on the subject in 1657, De ratiociniis in ludo aleae (\"On Reasoning in Games of Chance\"), which appeared as an appendix to Frans van Schooten's Exercitationes mathematicae. Some sources give 1656 for the Dutch original, so the dates differ slightly by edition. Huygens began from a postulate that the value of an uncertain prospect equals the sum that would buy the same prospect in a fair game. From this he derived that an equal chance of receiving either a or b is worth \\(\\frac{a+b}{2}\\), and he extended the rule to three or more outcomes and to multi-stage games. The treatise is the origin of the expected value concept in its explicit form, and it also covered a variant that later became known as the gambler's ruin problem.\n\nJakob Bernoulli spent much of his life on the same questions. His Ars Conjectandi was published posthumously in Basel in 1713 and contained the first version of the law of large numbers. For repeated trials with a fixed chance of success, he showed that the probability that the observed frequency lies within any chosen interval around the true chance approaches one as the number of trials grows. His proof gave no sense of how quickly this happens, and his bounds were loose: for a jar example aimed at estimating a probability of 0.6 within ±0.02 at odds of 1000:1, his method called for more than 25,000 observations. Bernoulli intended the \"art of conjecturing\" to guide action where certainty is impossible, and he hoped to extend the mathematics of games to civil, moral and economic questions.\n\nAbraham de Moivre sharpened this result. He first derived the normal approximation to the binomial distribution in a private printing of 1733, which he published in 1738, as an attempt to improve on Bernoulli's wide limits. He also tabulated the areas under the curve within one, two and three standard deviations of the mean. Pierre-Simon Laplace then generalised the result in his Théorie analytique des probabilités (1812), showing that the binomial tends toward the normal distribution as the number of trials grows, and the same work carried probability into astronomy, demography and the analysis of observational error. This line of development, from Huygens' expectations through Bernoulli's limit theorem to de Moivre and Laplace's approximations, is what later made the central limit theorem possible."
    },
    {
      "heading": "Axiomatic Foundations and Kolmogorov's Framework",
      "text": "Kolmogorov's Grundbegriffe der Wahrscheinlichkeitsrechnung (1933) replaced a body of informal rules with a short system of axioms. Glenn Shafer's historical study of the book notes that Kolmogorov covered the subject in roughly sixty pages. Probability space is built on three ingredients: a sample space Ω of possible outcomes, a σ-algebra \\(\\mathcal{F}\\) of events, and a function P that assigns each event a number. P must satisfy three axioms: non-negativity, \\(P(A)\\ge 0\\); normalization, \\(P(\\Omega)=1\\); and countable additivity, which states that for pairwise disjoint events \\(A_1, A_2, \\ldots\\),\n\n\\[P\\Big(\\bigcup_{i=1}^{\\infty} A_i\\Big)=\\sum_{i=1}^{\\infty} P(A_i).\\]\n\nFinite additivity is the special case of two or more events, and it follows from countable additivity once the family of events is closed under the relevant operations. The triple (Ω, \\(\\mathcal{F}\\), P) is the object that the rest of the theory, including random variables and distributions, is built on.\n\nCountable additivity was not Kolmogorov's invention. Émile Borel had introduced it into probability in 1909, and the following two decades saw many authors, Kolmogorov among them, explore its consequences. The Grundbegriffe merely rounded out the picture by explaining that nothing more was needed, as Shafer puts it. Steinhaus took a more explicit step in 1923: he axiomatized Borel's theory of denumerable probability on the model of Sierpiński's axioms for Lebesgue measure. Steinhaus concluded that the theory of probability for an infinite sequence of binary trials is isomorphic with the theory of Lebesgue measure. This showed that infinite coin-tossing sequences could be treated with the tools of measure theory.\n\nSeveral rival axiom systems were proposed before 1933. Sergei Bernstein's 1917 work showed that probability theory can be founded on qualitative axioms for numerical coefficients that measure the probabilities of propositions. Richard von Mises instead built his theory on the notion of a \"collective\", a sequence whose limiting frequencies are stable under selection, and he did not treat probability as a measure in Kolmogorov's sense. After years of debate, most probabilists were critical of von Mises' axiomatisation and preferred the simple set of axioms given in Kolmogorov's Grundbegriffe. Finite additivity alone was regarded as too weak. It cannot control the limits of increasing sequences of events, which the classical limit theorems require, and this is the main reason countable additivity was retained. Shafer also points to Kolmogorov's own work on the limit theorems as the source of the authority that let him declare the axioms sufficient.\n\nCountable additivity has its critics. Bruno de Finetti argued that it can conflict with intuitions about fair chance. Consider a lottery with countably many equally likely tickets: if every ticket has positive probability, the sum over infinitely many tickets exceeds one; if every ticket has probability zero, the sum is zero. Neither outcome fits a normalized total of one. Advocates of the axioms respond that such uniform distributions over infinite sets are simply not probability measures, and the debate has continued in the study of finitely additive and Bayesian alternatives.\n\nThe measure-theoretic viewpoint allowed discrete and continuous cases to be handled by a single definition. A discrete distribution assigns weights to countably many points, while a continuous one assigns probabilities through integration against a density with respect to Lebesgue measure. Both are special cases of a probability measure on the same σ-algebra, which is why the framework became the standard language of the field.\n\nKolmogorov also insisted that the axioms by themselves do not say what probability means in the world. The abstract theory is a mathematical structure, and its application requires a link to observable phenomena. For repeated experiments, the link is the stability of relative frequencies, expressed through the law of large numbers. The axioms do not decide which interpretation is correct, and the question of how a mathematically defined P corresponds to physical or epistemic chance remained open."
    },
    {
      "heading": "Elementary Rules and Combinatorial Methods",
      "text": "Complement, addition, and inclusion–exclusion\n\nFor an event \\(A\\) in a sample space \\(\\Omega\\), the complement rule gives \\(P(A^c)=1-P(A)\\). It is often the fastest route when the event of interest is \"at least one\" of something, because the complement is \"none.\"\n\nFor two events, the addition rule is \\(P(A\\cup B)=P(A)+P(B)-P(A\\cap B)\\). The subtraction removes the overlap counted twice. If \\(A\\) and \\(B\\) are mutually exclusive, \\(P(A\\cap B)=0\\) and the rule reduces to a plain sum. Mutual exclusivity is a strong condition and is often confused with independence, which is a different property (see below).\n\nThe inclusion–exclusion principle extends this to \\(n\\) events:\n\n\\[P\\Big(\\bigcup_{i=1}^{n} A_i\\Big)=\\sum_i P(A_i)-\\sum_{i<j} P(A_i\\cap A_j)+\\cdots+(-1)^{n+1}P\\Big(\\bigcap_{i=1}^{n} A_i\\Big).\\]\n\nA classic application is the hat-check problem. If \\(n\\) people return hats at random, the probability that at least one person gets their own hat is \\(1-\\tfrac{1}{2!}+\\tfrac{1}{3!}-\\cdots\\), which converges quickly to \\(1-1/e\\approx 0.632\\), regardless of how large \\(n\\) is.\n\nThe union bound\n\nThe union bound (Boole's inequality) states \\(P(\\bigcup_i A_i)\\le \\sum_i P(A_i)\\). It discards all overlap terms, so it is loose when events overlap heavily but needs no independence assumption. Suppose 1,000 independent tests each have a false-positive probability of 0.0001. The exact chance of at least one false positive is \\(1-(0.9999)^{1000}\\approx 0.095\\), while the union bound gives at most 0.1. The bound is useful precisely because it is simple and robust, which explains its frequent use in randomized algorithms and in proving concentration results.\n\nCounting in equiprobable finite spaces\n\nWhen every outcome in a finite sample space is equally likely, \\(P(A)=|A|/|\\Omega|\\), and the problem becomes one of counting. Ordered selections of \\(k\\) items from \\(n\\) distinct items number \\(n!/(n-k)!\\) (permutations). Unordered selections number \\(\\binom{n}{k}=\\dfrac{n!}{k!\\,(n-k)!}\\) (combinations). For example, the probability of exactly 5 heads in 10 fair coin tosses is \\(\\binom{10}{5}/2^{10}=252/1024\\approx 0.246\\); this is the binomial distribution in its simplest form. In the birthday problem, the probability that 23 people all have different birthdays is \\(\\prod_{i=0}^{22}(1-i/365)\\approx 0.493\\), so the probability of a shared birthday is about 0.507. The difficulty in such problems lies almost always in correctly identifying \\(\\Omega\\), not in the arithmetic.\n\nConditional probability\n\nFor \\(P(B)>0\\), the conditional probability of \\(A\\) given \\(B\\) is\n\n\\[P(A\\mid B)=\\frac{P(A\\cap B)}{P(B)}.\\]\n\nConceptually, conditioning restricts the sample space to the outcomes in \\(B\\) and rescales the probabilities so that they sum to one. The pitfalls are mostly about the event being conditioned on. First, \\(P(A\\mid B)\\) and \\(P(B\\mid A)\\) are generally different; confusing them underlies the prosecutor's fallacy. Second, the answer depends on how the conditioning information was obtained. In a family with two children, where at least one is a boy, the probability that both are boys is 1/3. If instead one is told that the first-born is a boy, the probability that both are boys is 1/2. The two statements describe different conditioning events, even though both mention a boy. Third, the elementary definition does not apply when \\(P(B)=0\\), which requires the more advanced machinery of conditional expectation and measure-theoretic constructions.\n\nIndependence and the law of total probability\n\nTwo events are independent when \\(P(A\\cap B)=P(A)P(B)\\). This is a property of the probability measure, not a logical consequence of the events' descriptions, and in applied work it is usually an assumption to be justified. For three or more events, pairwise independence (each pair satisfies the product rule) is weaker than mutual independence (every subcollection satisfies it). The standard counterexample uses two fair coin tosses: let \\(A\\) be \"first toss heads,\" \\(B\\) be \"second toss heads,\" and \\(C\\) be \"exactly one head.\" Each pair is independent, since each intersection has probability 1/4, but \\(P(A\\cap B\\cap C)=0\\neq 1/8=P(A)P(B)P(C)\\), so the three are not mutually independent.\n\nThe law of total probability connects conditional and unconditional quantities. If \\(B_1,\\dots,B_n\\) partition \\(\\Omega\\) with each \\(P(B_i)>0\\), then\n\n\\[P(A)=\\sum_{i=1}^{n}P(A\\mid B_i)\\,P(B_i).\\]\n\nFor instance, if machine 1 produces 60% of items with a 2% defect rate and machine 2 produces the remaining 40% with a 5% defect rate, the overall defect probability is \\(0.6\\times 0.02+0.4\\times 0.05=0.032\\). This identity is the denominator in Bayes' theorem, which is treated in the following section."
    },
    {
      "heading": "Bayes' Theorem and Conditional Reasoning",
      "text": "Statement and structure\n\nBayes' theorem relates a conditional probability to its reverse. For events \\(A\\) and \\(B\\) with \\(P(B)>0\\),\n\n\\[P(A\\mid B)=\\frac{P(B\\mid A)\\,P(A)}{P(B)}.\\]\n\nThe left side is the posterior probability of \\(A\\) after \\(B\\) has been observed. The numerator contains the likelihood \\(P(B\\mid A)\\), which measures how probable the evidence is if \\(A\\) holds, and the prior \\(P(A)\\), which records belief before the evidence arrives. The denominator is a normalizing constant. When \\(A\\) falls into a partition \\(A_1,\\dots,A_n\\), it can be expanded by the law of total probability as \\(P(B)=\\sum_i P(B\\mid A_i)P(A_i)\\). The formula then reads as \"posterior is proportional to likelihood times prior\", and the normalizing constant only ensures that the posterior values sum to one. The same structure underlies bayesian-inference, where \\(A\\) is a parameter or hypothesis, the prior is a prior distribution, and the updated result is a posterior distribution.\n\nBase rates and the interpretation of positive results\n\nThe theorem explains why a positive result from an accurate test can still be more likely to be wrong than right when the condition is rare. Suppose a disease has a prevalence of 1%, a test detects it with sensitivity 90%, and the false-positive rate is 5%. Among 1,000 people, about 9 true cases test positive and about 50 healthy people do too. The probability of disease given a positive result is therefore \\(9/59\\approx 15\\%\\), not the 90% that the sensitivity might suggest. The error comes from neglecting the base rate, and it is the textbook case of the base-rate fallacy. The same reasoning applies to screening programmes, eyewitness identification, and fraud detection, where the rarity of the target condition dominates the calculation.\n\nOdds form and the likelihood ratio\n\nDividing the theorem for \\(A\\) by the theorem for its complement removes the normalizing constant and yields the odds form:\n\n\\[\\frac{P(A\\mid B)}{P(\\neg A\\mid B)}=\\frac{P(A)}{P(\\neg A)}\\times\\frac{P(B\\mid A)}{P(B\\mid \\neg A)}.\\]\n\nPosterior odds equal prior odds multiplied by the likelihood ratio. In the example above, the prior odds are 1 to 99 and the likelihood ratio is \\(0.90/0.05=18\\), so the posterior odds are about 18/99, or roughly 0.18 to 1, which corresponds to the same probability of about 15%. The likelihood ratio is a convenient measure of how strongly the evidence favours one hypothesis over another, and it is the quantity used in the Bayes factor for comparing models. Because it does not depend on the prior, it can be reported separately from any particular belief about the base rate.\n\nHistorical background\n\nThe theorem bears the name of Thomas Bayes, an English Presbyterian minister, but his role was narrower than the name suggests. His essay was found among his papers after his death and was edited by his friend Richard Price, who read it to the Royal Society on 23 December 1763; it appeared the following year in the Philosophical Transactions. Bayes treated a specific problem: inferring the success probability of a binomial process from observed successes and failures, under a uniform prior. Price added an introduction that argued for the use of such inverse reasoning. Some historians note that a statement of the theorem appears earlier in David Hartley's 1749 book, where Hartley credits a friend, and that Stigler suggested Nicholas Saunderson as a possible source, although the evidence is not conclusive.\n\nPierre-Simon Laplace stated the rule in general form in his 1774 memoir Mémoire sur la probabilité des causes par les événements, where he called it the probability of causes. He reached the result independently and, according to accounts of his later work, only learned of Bayes's essay in 1781. Laplace applied the rule to population statistics, meteorology, geodesy, astronomy, and questions of testimony, and his 1812 Théorie analytique des probabilités gave it a systematic treatment. Many historians therefore regard the general theory as Laplace's contribution and the name as Bayes's.\n\nConditional reasoning in everyday judgment\n\nPsychological research from the early 1970s onward has tested whether people reason in line with the theorem. Daniel Kahneman and Amos Tversky showed that subjects tend to neglect base rates when given a diagnostic description, and that they often weigh the vividness of individual evidence more heavily than prior frequencies. Their studies of taxi-cab problems, in which the proportion of cab colours and the reliability of a witness were specified, showed answers that were typically far above the normatively correct value. Later work by Gerd Gigerenzer and Ulrich Hoffrage found that presenting the same information as natural frequencies (for example, \"10 of 1,000 people\") rather than as percentages substantially improved the rate of correct Bayesian answers. These findings are debated: some authors read them as a systematic bias in human reasoning, while others argue that the problem format, rather than the reasoning itself, is responsible for much of the error.\n\nThe relation between conditional reasoning and everyday judgment also depends on which probabilities people take as given. Conditional probability requires a stated conditioning event, and confusion between \\(P(A\\mid B)\\) and \\(P(B\\mid A)\\) is one of the most common errors in interpreting medical tests, legal evidence, and statistical significance."
    },
    {
      "heading": "Random Variables and Probability Distributions",
      "text": "Random variables as measurable functions\n\nA random variable is a function \\(X\\) from a sample space \\(\\Omega\\) to the real numbers \\(\\mathbb{R}\\). For \\(X\\) to be usable within a probability space \\((\\Omega, \\mathcal{F}, P)\\), it must be measurable: for every Borel set \\(B \\subseteq \\mathbb{R}\\), the preimage \\(\\{X \\in B\\} = \\{\\omega : X(\\omega) \\in B\\}\\) must belong to the σ-algebra \\(\\mathcal{F}\\), so that its probability is defined. Taking \\(B = (-\\infty, x]\\) gives the events \\(\\{X \\le x\\}\\), which are the basic building blocks of everything that follows.\n\nThrough \\(X\\), the measure \\(P\\) on \\(\\Omega\\) induces a measure on \\(\\mathbb{R}\\), the probability distribution of \\(X\\). Questions about \\(X\\) are, in practice, questions about this induced distribution. For example, if two coins are tossed and \\(X\\) counts the heads, \\(\\Omega\\) has four equally likely outcomes, while \\(X\\) takes the values 0, 1 and 2 with probabilities 1/4, 1/2 and 1/4. Different random variables defined on different spaces can share the same distribution, and in that sense only the distribution matters for many purposes.\n\nDiscrete and continuous distributions\n\nWhen \\(X\\) takes countably many values, its distribution is described by a probability mass function \\(p(x) = P(X = x)\\). The values are nonnegative and sum to 1. The binomial case is a standard example:\n\n\\[p(k) = \\binom{n}{k} p^k (1-p)^{n-k}, \\quad k = 0, 1, \\dots, n,\\]\n\nwhich is the binomial distribution.\n\nWhen the distribution is absolutely continuous, it is described by a probability density function \\(f \\ge 0\\) satisfying\n\n\\[P(a \\le X \\le b) = \\int_a^b f(x)\\,dx, \\qquad \\int_{-\\infty}^{\\infty} f(x)\\,dx = 1.\\]\n\nFor such variables every single point has probability zero, so the density is not itself a probability. Instead, \\(f(x)\\,h\\) approximates \\(P(x \\le X \\le x+h)\\) for small \\(h\\), which is why the density is better read as a concentration of probability per unit length. The normal distribution is the best-known continuous example.\n\nNot every distribution falls into one of these two types. The Lebesgue decomposition, which relies on measure theory and Lebesgue measure, splits a general distribution into an absolutely continuous part, a discrete part, and a singular part with no density that is nonetheless not discrete.\n\nCumulative distribution functions\n\nThe cumulative distribution function (CDF) is defined for every real-valued random variable as \\(F(x) = P(X \\le x)\\). It is non-decreasing, right-continuous, and tends to 0 as \\(x \\to -\\infty\\) and to 1 as \\(x \\to +\\infty\\). Conversely, any function with these three properties is the CDF of some distribution, so \\(F\\) determines the distribution completely. This universality is the reason the CDF serves as the common language for discrete, continuous and mixed cases.\n\nThe CDF also encodes the probabilities directly. We have \\(P(a < X \\le b) = F(b) - F(a)\\), and the probability of a point \\(x\\) equals the size of the jump of \\(F\\) at \\(x\\). For a discrete variable, \\(F\\) is a step function; for an absolutely continuous one, \\(F\\) is continuous and \\(F'(x) = f(x)\\) wherever the derivative exists.\n\nJoint, marginal and conditional distributions\n\nFor two random variables \\(X\\) and \\(Y\\) on the same space, the joint CDF is \\(F(x,y) = P(X \\le x, Y \\le y)\\). In the discrete case the joint mass function is \\(p(x,y)\\), and in the continuous case the joint density is \\(f(x,y)\\). The marginal distributions are obtained by summing or integrating out the other variable:\n\n\\[p_X(x) = \\sum_y p(x,y), \\qquad f_X(x) = \\int f(x,y)\\,dy.\\]\n\nThe conditional distribution of \\(Y\\) given \\(X = x\\) is \\(p(y \\mid x) = p(x,y)/p_X(x)\\) in the discrete case, and \\(f(y \\mid x) = f(x,y)/f_X(x)\\) in the continuous case where \\(f_X(x) > 0\\). This is the distributional counterpart of conditional probability; conditioning on an event of probability zero in the continuous setting requires a more careful construction.\n\nRandom variables are independent when their joint distribution factorizes: \\(F(x,y) = F_X(x)F_Y(y)\\) for all \\(x, y\\), or equivalently \\(p(x,y) = p_X(x)p_Y(y)\\) or \\(f(x,y) = f_X(x)f_Y(y)\\). Independence is a much stronger requirement than zero covariance. For instance, if \\(X\\) is standard normal and \\(Y = X^2\\), then \\(X\\) and \\(Y\\) are uncorrelated yet clearly dependent.\n\nTransformations of random variables\n\nIf \\(Y = g(X)\\), the distribution of \\(Y\\) is determined by that of \\(X\\) and by \\(g\\). For discrete variables, the mass at \\(y\\) is the sum of the masses of all \\(x\\) with \\(g(x) = y\\). For continuous variables and a differentiable, strictly monotone \\(g\\) with inverse \\(g^{-1}\\), the change-of-variables formula gives\n\n\\[f_Y(y) = f_X\\big(g^{-1}(y)\\big)\\,\\left|\\frac{d}{dy}g^{-1}(y)\\right|.\\]\n\nFor example, if \\(X\\) is uniform on \\((0,1)\\) and \\(Y = -\\ln X\\), then \\(f_Y(y) = e^{-y}\\) for \\(y > 0\\), which is the exponential distribution with rate 1.\n\nWhen \\(g\\) is not monotone, the contributions from all preimages are added:\n\n\\[f_Y(y) = \\sum_{i} \\frac{f_X(x_i)}{|g'(x_i)|}, \\qquad g(x_i) = y.\\]\n\nIf \\(X\\) is standard normal and \\(Y = X^2\\), each \\(y > 0\\) has two preimages \\(\\pm\\sqrt{y}\\), and the formula yields \\(f_Y(y) = (2\\pi y)^{-1/2} e^{-y/2}\\), the chi-squared distribution with one degree of freedom. In several dimensions the same idea uses the absolute value of the Jacobian determinant in place of \\(|g'|\\)."
    },
    {
      "heading": "Expectation, Variance, and Probability Inequalities",
      "text": "Expected value\n\nFor a discrete random variable \\(X\\) taking values \\(x_i\\) with probabilities \\(p_i\\), the expected value is \\(E[X]=\\sum_i x_i p_i\\). For a continuous variable with density \\(f\\), it is \\(E[X]=\\int x f(x)\\,dx\\). In both cases the expectation is a weighted average of the possible values, each weighted by its probability. A fair six-sided die has expectation 3.5, a value the die never shows. Expectation is linear: \\(E[aX+bY]=aE[X]+bE[Y]\\) holds whether or not \\(X\\) and \\(Y\\) are independent. This property makes many calculations simple; for example, a binomial count with \\(n\\) trials and success probability \\(p\\) has mean \\(np\\), obtained by summing \\(n\\) Bernoulli means. Expectation also locates the centre of a distribution. Among all constants \\(c\\), the quantity \\(E[(X-c)^2]\\) is smallest when \\(c=E[X]\\). Some distributions have no finite expectation. The Cauchy distribution is the standard example, because the integral defining its mean does not converge absolutely.\n\nVariance, covariance, and correlation\n\nThe variance \\(\\operatorname{Var}(X)=E[(X-\\mu)^2]\\), where \\(\\mu=E[X]\\), measures spread around the mean. It can also be computed as \\(E[X^2]-\\mu^2\\). Its square root, the standard deviation, has the same units as \\(X\\). Scaling multiplies variance by the square of the scale factor, \\(\\operatorname{Var}(aX+b)=a^2\\operatorname{Var}(X)\\), so adding a constant leaves spread unchanged. Unlike expectation, variance is not linear. Instead, \\(\\operatorname{Var}(X+Y)=\\operatorname{Var}(X)+\\operatorname{Var}(Y)+2\\operatorname{Cov}(X,Y)\\).\n\nFor two variables, the covariance is \\(\\operatorname{Cov}(X,Y)=E[(X-E[X])(Y-E[Y])]\\). The correlation is \\(\\rho=\\operatorname{Cov}(X,Y)/(\\sigma_X\\sigma_Y)\\), a dimensionless quantity between −1 and 1. Independent variables have zero covariance, but the converse fails. If \\(X\\) is standard normal and \\(Y=X^2\\), then \\(\\operatorname{Cov}(X,Y)=E[X^3]=0\\), yet \\(Y\\) is completely determined by \\(X\\). Zero correlation therefore rules out only linear association, not dependence in general.\n\nHigher moments and the moment generating function\n\nThe \\(k\\)-th moment is \\(E[X^k]\\), and the \\(k\\)-th central moment is \\(E[(X-\\mu)^k]\\). Standardising the third central moment gives skewness, \\(E[((X-\\mu)/\\sigma)^3]\\), which measures asymmetry. The standardised fourth moment gives kurtosis, \\(E[((X-\\mu)/\\sigma)^4]\\). A normal distribution has kurtosis 3, and \"excess kurtosis\" subtracts this value to compare tail weight with the normal case.\n\nThe moment generating function \\(M_X(t)=E[e^{tX}]\\) collects all moments in one function. When it exists on an interval around zero, the \\(k\\)-th moment equals the \\(k\\)-th derivative of \\(M_X\\) at 0, and the function determines the distribution uniquely. For independent variables, \\(M_{X+Y}=M_X M_Y\\), which is why it is useful for identifying distributions under addition. A normal variable with mean \\(\\mu\\) and variance \\(\\sigma^2\\) has \\(M(t)=\\exp(\\mu t+\\sigma^2 t^2/2)\\). Heavy-tailed families such as the Cauchy have no moment generating function, and the characteristic function \\(E[e^{itX}]\\), which always exists, is used in their place.\n\nMarkov's and Chebyshev's inequalities\n\nMarkov's inequality states that for a nonnegative random variable \\(X\\) and any \\(a>0\\), \\(P(X\\ge a)\\le E[X]/a\\). It requires only a finite mean, and it follows from the observation that \\(a\\mathbf{1}\\{X\\ge a\\}\\le X\\). Chebyshev's inequality follows by applying Markov's inequality to \\((X-\\mu)^2\\) with threshold \\(k^2\\sigma^2\\), giving \\(P(|X-\\mu|\\ge k\\sigma)\\le 1/k^2\\). At \\(k=2\\), the bound is 0.25 for any distribution with finite variance. For a normal distribution the true two-sided tail beyond two standard deviations is about 0.046, so the bound is loose for well-behaved distributions. Its strength lies in requiring no knowledge of the shape, and it is the basic tool in proofs of the weak law of large numbers. Both bounds are attained by some distributions, so they cannot be improved without further assumptions.\n\nJensen's inequality\n\nJensen's inequality states that if \\(\\varphi\\) is convex, then \\(\\varphi(E[X])\\le E[\\varphi(X)]\\), provided the expectations exist. For concave \\(\\varphi\\) the inequality reverses. Geometrically, a convex function lies below its chords, so its value at the mean does not exceed the average of its values. Several familiar facts follow directly. The expected absolute value satisfies \\(E|X|\\ge|E[X]|\\). The inequality \\(E[X^2]\\ge(E[X])^2\\) is equivalent to the nonnegativity of variance. For positive \\(X\\), \\(E[1/X]\\ge 1/E[X]\\), and \\(E[e^{tX}]\\ge e^{tE[X]}\\). Equality holds only when \\(\\varphi\\) is linear over the range of \\(X\\) or when \\(X\\) is almost surely constant."
    },
    {
      "heading": "Standard Families of Distributions",
      "text": "Most probability models used in practice are drawn from a small set of parametric families. Each family is tied to a recurring kind of situation, so knowing what a family describes helps decide whether it fits a given problem before any data are fitted.\n\nA Bernoulli trial has two outcomes, coded 1 (success) and 0 (failure), with success probability \\(p\\). Repeating such a trial \\(n\\) times, independently and with the same \\(p\\), and counting successes gives the binomial distribution, with mass function \\(\\binom{n}{k}p^k(1-p)^{n-k}\\), mean \\(np\\) and variance \\(np(1-p)\\). If the draws are made without replacement from a finite population of \\(N\\) items containing \\(K\\) successes, the count instead follows the hypergeometric distribution, with probabilities \\(\\binom{K}{k}\\binom{N-K}{n-k}/\\binom{N}{n}\\). When the sample is small relative to the population, the two families give nearly identical results, which is why the binomial is routinely used for small sampling fractions.\n\nFor rare events, the Poisson distribution arises as a limit of the binomial: letting \\(n\\to\\infty\\) and \\(p\\to 0\\) while \\(np=\\lambda\\) stays fixed yields \\(P(k)=e^{-\\lambda}\\lambda^k/k!\\), and both its mean and its variance equal \\(\\lambda\\). Its waiting-time counterpart is the exponential distribution, with density \\(\\lambda e^{-\\lambda x}\\) for \\(x\\ge 0\\). The exponential is memoryless: the probability of waiting a further interval does not depend on how long one has already waited. Successive gaps between events of a Poisson process are exponentially distributed, which links the two families directly.\n\nAmong continuous families, the uniform distribution on \\([a,b]\\) has constant density \\(1/(b-a)\\) and serves as the continuous analogue of equal likelihood, often as an uninformative starting point. The gamma distribution with shape \\(\\alpha\\) and rate \\(\\beta\\) is the law of a sum of \\(\\alpha\\) independent exponentials when \\(\\alpha\\) is an integer, and it covers the exponential (\\(\\alpha=1\\)) and chi-squared cases; it is used for positive quantities such as waiting times and claim sizes. The beta distribution on \\([0,1]\\) models proportions and unknown probabilities. Gamma and beta distributions are also conjugate priors: a gamma prior for a Poisson rate and a beta prior for a binomial \\(p\\) yield posteriors of the same family, so a \\(\\mathrm{Beta}(\\alpha,\\beta)\\) prior updated with \\(k\\) successes in \\(n\\) trials becomes \\(\\mathrm{Beta}(\\alpha+k,\\beta+n-k)\\). The uniform distribution is the special case \\(\\mathrm{Beta}(1,1)\\).\n\nThe normal distribution is the most widely used family, specified by a mean \\(\\mu\\) and variance \\(\\sigma^2\\). It has several characterizations. Among all distributions with a given mean and variance, it has maximum entropy; it is closed under linear combinations of independent normal variables; and, under classical conditions, a random sample from a distribution has its sample mean and sample variance statistically independent only when that distribution is normal. Its prominence in practice is largely explained by the central limit theorem. The multivariate normal distribution is determined by a mean vector \\(\\mu\\) and a covariance matrix \\(\\Sigma\\), with density proportional to \\(\\exp\\left(-\\tfrac12(x-\\mu)^{T}\\Sigma^{-1}(x-\\mu)\\right)\\). Linear transformations, marginals and conditional distributions of a multivariate normal vector are again normal, and within such a vector, zero covariance between two components implies their independence, a property that fails for general distributions.\n\nNormal and exponential tails decay rapidly, and these are called light-tailed families: their moments of all orders are finite, and their moment generating functions exist in a neighbourhood of zero. Heavy-tailed families behave differently. The Cauchy distribution, with density \\(1/(\\pi(1+x^2))\\), has no finite mean, and the average of \\(n\\) independent standard Cauchy variables has exactly the same distribution as a single one, so averaging does not reduce spread and the law of large numbers does not apply. Power-law families such as the Pareto distribution have finite moments only up to some order, determined by the tail exponent. The distinction matters in risk modelling. Losses in financial markets, insurance claims and some natural hazards show heavy tails, and fitting a normal model to such data can greatly underestimate the probability of extreme outcomes, while the sample variance and other moment-based summaries may be unstable."
    },
    {
      "heading": "Limit Theorems and Modes of Convergence",
      "text": "Laws of large numbers\n\nLet \\(X_1, X_2, \\dots\\) be independent and identically distributed (i.i.d.) random variables with finite mean \\(\\mu\\), and let \\(\\bar X_n = \\frac{1}{n}\\sum_{i=1}^n X_i\\). The weak law of large numbers states that for every \\(\\varepsilon > 0\\), \\(P(|\\bar X_n - \\mu| > \\varepsilon) \\to 0\\), so \\(\\bar X_n\\) converges to \\(\\mu\\) in probability. Khinchin proved this under the finite-mean assumption alone. The strong law, established by Kolmogorov in 1930 for the i.i.d. finite-mean case, asserts that \\(P(\\bar X_n \\to \\mu) = 1\\), that is, convergence almost surely. Borel's 1909 result on the digits of fair coin tosses was an early special case. The distinction matters: the weak law concerns the probability of a large deviation at each fixed \\(n\\), while the strong law concerns the behaviour of the whole sequence of averages along a single realization.\n\nCentral limit theorem\n\nIf in addition \\(\\sigma^2 = \\operatorname{Var}(X_i) \\in (0, \\infty)\\), the Lindeberg–Lévy form of the central limit theorem gives \\[\\frac{\\sqrt{n}\\,(\\bar X_n - \\mu)}{\\sigma} \\xrightarrow{\\ d\\ } N(0,1).\\] The limit depends only on the mean and variance of the summands, which explains why the normal distribution appears so widely: standardized sums of many small, roughly independent contributions forget the detailed shape of each one. The theorem extends to independent but non-identical summands. The Lyapunov condition requires a moment slightly above the second, while the Lindeberg–Feller theorem gives a necessary and sufficient condition, the Lindeberg condition, which essentially says that no single term dominates the total variance. The multivariate version yields a multivariate normal limit with the covariance matrix of the summands. The Berry–Esseen theorem bounds the approximation error at order \\(n^{-1/2}\\), scaled by the third absolute moment, which explains why strongly skewed distributions need larger samples before the normal approximation is accurate.\n\nModes of convergence\n\nFor random variables \\(X_n\\) and \\(X\\), almost sure convergence means \\(P(\\lim_n X_n = X) = 1\\). Convergence in probability means \\(P(|X_n - X| > \\varepsilon) \\to 0\\) for every \\(\\varepsilon > 0\\). Convergence in distribution means that the distribution functions converge at every continuity point of the limit; it concerns only distributions, and the central limit theorem is of this type. Convergence in \\(L^p\\) means \\(E|X_n - X|^p \\to 0\\) (see L^p convergence). Almost sure convergence and \\(L^p\\) convergence each imply convergence in probability, which in turn implies convergence in distribution. Neither of the first two implies the other. On \\([0,1]\\) with Lebesgue measure, \\(X_n = n\\,\\mathbf{1}_{(0,1/n)}\\) tends to 0 almost surely, yet \\(E X_n = 1\\) for all \\(n\\). Conversely, a \"typewriter\" sequence of indicators of shrinking intervals sweeping across \\([0,1]\\) converges in probability to 0 while failing to converge at any point. Convergence in distribution to a constant is equivalent to convergence in probability, and every sequence converging in probability has an almost surely convergent subsequence.\n\nBorel–Cantelli lemmas\n\nFor events \\(A_n\\), let \\(\\{A_n \\text{ i.o.}\\}\\) denote the event that infinitely many of them occur. The first Borel–Cantelli lemma states that if \\(\\sum_n P(A_n) < \\infty\\), then \\(P(A_n \\text{ i.o.}) = 0\\), with no independence assumption. The second lemma states that if the events are independent and \\(\\sum_n P(A_n) = \\infty\\), then \\(P(A_n \\text{ i.o.}) = 1\\). For independent events with \\(P(A_n) = 1/n\\), infinitely many occur almost surely; with \\(P(A_n) = 1/n^2\\), only finitely many do. These lemmas are a standard tool for deciding almost sure questions and for proving strong laws from moment bounds.\n\nLimits of the theorems\n\nThe laws and the central limit theorem depend on their hypotheses. A standard Cauchy variable has no finite mean, and the average of \\(n\\) i.i.d. Cauchy variables has the same Cauchy distribution for every \\(n\\), so neither law applies. When the mean is finite but the variance is infinite, the weak law still holds, but the normal limit fails; suitably rescaled sums instead converge to stable laws with index \\(\\alpha \\in (0,2)\\), with scaling \\(n^{1/\\alpha}\\). Dependence can also break convergence: if \\(X_n = X_1\\) for all \\(n\\), each marginal distribution is unchanged, yet \\(\\bar X_n = X_1\\) never approaches \\(\\mu\\). Weaker forms of the theorems do hold for dependent sequences, including the ergodic theorem for Markov chains and central limit theorems under mixing conditions."
    },
    {
      "heading": "Stochastic Processes",
      "text": "A stochastic process is a family of random variables \\(\\{X_t : t \\in T\\}\\) indexed by time, which describes how a random quantity evolves. The simplest and most widely used case is the discrete-time Markov chain, in which the next state depends only on the current state. For a countable state space, the Markov property is \\(P(X_{n+1}=j \\mid X_n=i, \\dots, X_0=i_0) = P(X_{n+1}=j \\mid X_n=i) = p_{ij}\\). The numbers \\(p_{ij}\\) form a transition matrix \\(P\\) whose rows sum to 1, and the \\(n\\)-step transition probabilities are the entries of \\(P^n\\). A distribution \\(\\pi\\) satisfying \\(\\pi P = \\pi\\) is a stationary distribution. For an irreducible, positive recurrent chain, such a distribution exists and is unique, and the ergodic theorem states that time averages along a single trajectory converge to averages under \\(\\pi\\), almost surely. This is the chain-level analogue of the law of large numbers. Andrey Markov introduced these chains in 1906 while studying vowel sequences in Pushkin's verse, and the theory now underpins Markov chain Monte Carlo and hidden Markov models.\n\nA random walk is the cumulative sum \\(S_n = \\xi_1 + \\cdots + \\xi_n\\) of independent, identically distributed steps. In the simplest case each step is \\(\\pm 1\\). Such walks are recurrent in one and two dimensions, so the walker returns to the origin with probability 1, but in three or more dimensions there is a positive probability of never returning. The gambler's ruin problem asks for the chance that a player starting with capital \\(i\\) reaches a target \\(N\\) before going broke, when each unit bet is won with probability \\(p\\) and lost with probability \\(q = 1-p\\). For \\(p \\neq q\\), the probability of reaching \\(N\\) first is \\(\\frac{1-(q/p)^i}{1-(q/p)^N}\\), and for a fair game (\\(p = 1/2\\)) it is \\(i/N\\). Even with a fair game, the ruin probability approaches 1 as the opponent's capital grows without bound, which is one reason the problem is central to discussions of gambler's ruin in finance and betting.\n\nThe Poisson process models the number of random events \\(N(t)\\) that occur up to time \\(t\\). Its defining features are independent increments and, for each interval of length \\(s\\), a count that follows a Poisson distribution with mean \\(\\lambda s\\). Equivalently, the waiting times between successive events are independent exponential variables with rate \\(\\lambda\\), and the exponential distribution is memoryless. This makes the process the basic building block of queueing models, such as the M/M/1 queue, where arrivals and service completions are both Poisson-driven. In reliability engineering, it describes the occurrence of component failures and repairs when failures arise independently at a constant rate.\n\nBrownian motion (or the Wiener process) is the continuous-time analogue of the random walk. A standard version \\(W_t\\) starts at zero, has continuous sample paths, and has independent Gaussian increments with \\(W_t - W_s \\sim N(0, t-s)\\). Louis Bachelier used a similar model in 1900 to describe price fluctuations, and Albert Einstein's 1905 analysis of particle diffusion gave it a physical basis, which Jean Perrin's experiments subsequently confirmed. Norbert Wiener gave a rigorous construction in the 1920s. The paths are nowhere differentiable, so the calculus needed to work with them is the stochastic calculus of Itô. In mathematical finance, geometric Brownian motion is the standard model for asset prices and underlies the Black–Scholes option pricing framework, while in physics the same process describes diffusion and thermal motion.\n\nA martingale formalizes the idea of a fair game: the conditional expectation of the next value, given the past, equals the current value, \\(E[X_{n+1} \\mid \\mathcal{F}_n] = X_n\\). Martingales are closely tied to conditional expectation and to the filtration \\(\\mathcal{F}_n\\) that records the information available up to time \\(n\\). Doob's optional stopping theorem states that, under suitable boundedness or integrability conditions, the expected value at a stopping time equals the initial expected value. Applied to a suitable martingale, this gives a short derivation of gambler's ruin probabilities. Martingale convergence theorems state that a nonnegative supermartingale converges almost surely, and that an \\(L^1\\)-bounded martingale converges almost surely to a limit. These results are fundamental in both probability theory and mathematical finance, where the absence of arbitrage is expressed through the existence of an equivalent martingale measure."
    },
    {
      "heading": "Interpretations of Probability",
      "text": "The formal theory of probability, built on Kolmogorov's axioms, specifies how probability measures behave but does not say what a probability number refers to in the world. Several interpretations fill this gap. They differ mainly in what they take probabilities to be about: symmetries of a setup, long-run frequencies, physical propensities, or rational degrees of belief.\n\nClassical interpretation\n\nThe oldest account treats probability as a ratio of equally possible cases: the number of favourable cases divided by the total number of equally possible cases. It rests on the principle of indifference (also called the principle of insufficient reason), which holds that outcomes with no reason to be preferred over one another should receive equal probabilities. This works well for fair coins, dice and well-shuffled cards, where symmetry is evident. Its limits are serious. Most real problems offer no natural partition into equal cases. The principle also gives inconsistent results under reparametrization: ignorance about a quantity x can be expressed as uniform in x or uniform in x², and the two choices assign different probabilities to the same event. Bertrand's paradox shows that a simple geometric question about a random chord yields several answers, depending on the measure chosen. The classical view therefore survives mainly as a special case rather than as a general definition.\n\nFrequentist interpretation\n\nFrequentism identifies probability with the limiting relative frequency of an outcome in an indefinitely long series of repetitions. Richard von Mises made this precise in 1919 with the concept of the collective: an infinite sequence in which relative frequencies converge and stay stable under any admissible selection of subsequences, which he took to formalize randomness. Frequentist ideas underpin much of twentieth-century statistics, including the methods associated with Ronald Fisher, Jerzy Neyman and Egon Pearson. The interpretation is valued for its objectivity and its direct link to observed data. Its weaknesses are that infinite sequences are idealizations, that it assigns no probability to a unique event such as a particular election result, and that single-case questions must be pushed into a reference class whose choice can alter the answer.\n\nPropensity interpretations\n\nPropensity views hold that probability is a physical tendency or disposition of a particular experimental arrangement to produce an outcome. On this account it is a property of the setup rather than of a sequence of trials. Karl Popper developed the proposal in the late 1950s, partly to allow single-case probabilities in quantum physics and partly to avoid the infinite-sequence problem. Other philosophers and physicists have developed related versions, some of which connect propensity to the chance dispositions of quantum systems. Critics point out that propensities are not directly observable and must be tested through frequencies. They also note that conditional propensities raise technical puzzles, such as Humphreys' paradox. The propensity view thus sits between frequentist testability and a physical account of chance.\n\nSubjective and Bayesian interpretations\n\nSubjectivists take probability to be a degree of belief held by a rational agent. Bayesian approaches adopt this reading. Frank Ramsey in 1926 and Bruno de Finetti in the 1930s argued that degrees of belief can be measured through betting prices and that a coherent agent's beliefs must satisfy the probability axioms. De Finetti's coherence criterion requires that no set of bets, each fair by the agent's own prices, guarantees a loss whatever the outcome. The Dutch book argument makes this precise by claiming that incoherent agents can be exploited in this way. Updating on new evidence then follows from the same reasoning and links subjective probability to Bayes' theorem. De Finetti's representation theorem for exchangeable sequences also shows that beliefs of this kind can behave as if they were drawn from unknown frequencies, which connects the subjective view to frequentist practice. Critics question whether betting prices faithfully reveal beliefs, whether the Dutch book argument establishes more than consistency, and how strongly the choice of prior affects conclusions.\n\nRelation to the formalism and the continuing debate\n\nAll these interpretations accept the same formal framework: a probability-space with non-negativity, normalization and countable additivity. Their disagreement concerns the link between the abstract measure and the world, not the calculus itself. In practice, statisticians often combine the views. Frequency-based models are used to design experiments and evaluate procedures, while Bayesian updating expresses changing belief about a hypothesis. Physics adds a further question, since the Born rule in quantum theory raises the issue of whether probabilities of single measurements are fundamental or reflect incomplete knowledge. The debate continues over which interpretation best fits scientific practice. Each captures some uses well: repeatable processes suit frequencies, unrepeatable events suit beliefs, and physical setups suit propensities. None has won general acceptance."
    },
    {
      "heading": "Probability in Statistical Inference",
      "text": "Estimators and their sampling behaviour\n\nStatistical inference runs probabilistic reasoning backwards, from an observed sample to an unknown parameter θ. An estimator is a rule that turns data into a guess of θ, and its quality is judged by its sampling-distribution, the distribution of its values across hypothetical repeated samples. An estimator is unbiased when its expected value equals the parameter. The sample mean meets this condition for a population mean, but the sample variance computed with divisor n falls short of σ² by a factor of (n−1)/n, which is why the divisor n−1 is standard. Unbiasedness alone does not settle the matter, since an estimator can have small bias and large variance. The bias-variance-tradeoff describes how reducing one often inflates the other. Consistency requires that the estimate converge to the true value as the sample grows, a property that rests on the law-of-large-numbers.\n\nMaximum likelihood and Bayesian updating\n\nFor independent observations x₁, …, xₙ from a model with density f(x; θ), the likelihood-function is \\(L(\\theta) = \\prod_{i=1}^{n} f(x_i;\\theta)\\), read as a function of θ with the data held fixed. maximum-likelihood-estimation selects the θ that maximizes it. For n Bernoulli trials with k successes, the maximizer is k/n. Under standard regularity conditions the resulting estimators are consistent and asymptotically normal, and ronald-fisher made likelihood a central tool of estimation in the 1920s.\n\nThe Bayesian approach treats θ itself as random. A prior-distribution π(θ) encodes beliefs before the data, and bayes-theorem yields the posterior-distribution \\(\\pi(\\theta \\mid x) \\propto L(\\theta)\\,\\pi(\\theta)\\). With a uniform prior, 7 successes in 10 trials give a posterior \\(\\text{Beta}(8,4)\\) (beta-distribution), with mean 2/3. A bayesian-inference credible interval is an interval containing a stated share of posterior probability, such as 95%. Each posterior becomes the prior for the next batch of data, which is the sense in which the method is sequential updating.\n\nHypothesis tests, p-values, and interval estimates\n\nA test begins with a null-hypothesis H₀ and a test statistic. The p-value is the probability, computed under H₀, of a statistic at least as extreme as the one observed. It is not the probability that H₀ is true. A 2016 statement by the American Statistical Association makes this point directly, listing it among six principles on the use of p-values. p-values do not measure the probability that the studied hypothesis is true, or the probability that the data were produced by random chance alone.\n\nDecisions carry two kinds of error, described in type-i-and-type-ii-errors. A type I error rejects a true null, and its probability is the significance level α. A type II error fails to reject a false null, with probability β. The statistical-power of a test is 1−β. A 95% confidence-interval is a procedure that, across repeated samples, covers the true value in about 95% of cases; it does not assign a 95% probability to the particular interval already computed. The framework that treats p-values as graded evidence associated with Fisher and the decision-oriented framework associated with Neyman and Pearson are often combined in applied practice, and the combination is one source of confusion.\n\nPower and multiple comparisons\n\nPower depends on the effect size, the sample size, the variability of the data, and α. Underpowered studies miss real effects often, and when they do produce significant results, the estimated effects tend to be inflated. Power analysis is therefore carried out before data collection to choose a sample size. Multiple comparisons create a separate problem. If m independent tests are each run at α = 0.05 and every null hypothesis is true, the probability of at least one false positive is \\(1 - 0.95^{m}\\), about 0.64 when m = 20. Procedures such as the Bonferroni correction, which tests each hypothesis at α/m, control the family-wise error rate at the cost of power.\n\nThe debate over misuse and reproducibility\n\nConcern about replication has made significance testing a public controversy. The 2016 ASA statement was written to improve the conduct and interpretation of quantitative science and inform the growing emphasis on reproducibility of science research. It stressed that pragmatic considerations often require binary, \"yes-no\" decisions, but this does not mean that p-values alone can ensure that a decision is correct or incorrect. A 2019 editorial went further and called for dropping the phrase \"statistically significant\" and the threshold p < 0.05. In response, the ASA's president convened a task force in 2019. Its report argued that the use of P-values and significance testing, properly applied and interpreted, are important tools that should not be abandoned. The report also held that thresholds should be set explicitly from study goals and the consequences of wrong decisions. Proposed remedies in the broader debate include reporting effect sizes and confidence intervals, using Bayesian measures, and preregistering analyses. The disagreement over which of these should replace or supplement the p-value remains open."
    },
    {
      "heading": "Physics: Statistical Mechanics and Quantum Theory",
      "text": "Statistical mechanics\n\nStatistical mechanics describes systems with enormous numbers of particles by assigning probabilities to their microscopic configurations instead of tracking each particle. For a system in thermal equilibrium, the basic probability law is the Boltzmann distribution. Each state with energy \\(E_i\\) receives a probability proportional to an exponential of \\(-E_i/k_BT\\), divided by a normalizing sum over all states:\n\n\\[p_i = \\frac{e^{-E_i/k_BT}}{Z}, \\qquad Z = \\sum_j e^{-E_j/k_BT}\\]\n\nHere \\(k_B\\) is the Boltzmann constant, \\(T\\) is the absolute temperature, and \\(Z\\) is the partition function, which makes the probabilities sum to one. Boltzmann first formulated this distribution in 1868 while studying the statistical mechanics of gases in thermal equilibrium.\n\nThe distribution has several consequences that follow directly from its form. Taking the ratio of two state probabilities cancels \\(Z\\), so only the energy difference matters: the ratio of the probabilities of two states is an exponential of the negative energy difference divided by \\(k_BT\\). Temperature controls how sharply the probabilities differ. At high temperature, states with different finite energies have more nearly equal probabilities, while at low temperature, lower-energy states receive exponentially greater probability.\n\nThe partition function links microscopic probabilities to macroscopic thermodynamics. Free energy is obtained as \\(F = -k_BT\\ln Z\\), and energy and other response quantities follow from derivatives of \\(\\ln Z\\). Entropy is given by the Gibbs form \\(S = -k_B\\sum_i p_i\\ln p_i\\), which reduces to Boltzmann's \\(S = k_B\\ln W\\) when all \\(W\\) accessible microstates are equally probable. In quantum statistical mechanics the classical weights are replaced by the density operator \\(\\rho = e^{-\\beta H}/\\mathrm{Tr}\\,e^{-\\beta H}\\). Identical particles require Bose–Einstein or Fermi–Dirac occupation statistics, and classical counting fails for phenomena such as blackbody radiation and electron gases in metals.\n\nQuantum theory\n\nIn quantum mechanics, probability enters as a rule for measurement rather than as a consequence of ignorance about a hidden state. The Born rule was introduced by Max Born in 1926, in a paper on the quantum mechanics of collision processes. For a normalized eigenstate \\(\\varphi\\) of a system in state \\(\\psi\\), the probability of finding the system in that eigenstate is \\(|\\langle\\varphi|\\psi\\rangle|^2\\), and for position measurements the probability density is \\(|\\psi(x)|^2\\), normalized so that its integral over all space equals one.\n\nBorn's own thinking was not immediately settled. At first he believed the amplitude of the wave function itself gave the probability, and he soon concluded that the squared modulus was the correct quantity. Because the rule is applied to complex amplitudes before squaring, the resulting probabilities contain cross terms. These produce interference patterns, such as those seen in the double-slit experiment, that have no classical counterpart.\n\nHeisenberg's uncertainty principle is a statistical statement about the spread of measured values. Werner Heisenberg's 1927 analysis argued that position and momentum cannot both be determined precisely at the same time. In its standard modern form for a pair of conjugate observables, the standard deviations of position and momentum satisfy \\(\\sigma_x\\sigma_p \\ge \\hbar/2\\), where \\(\\hbar\\) is the reduced Planck constant. The bound concerns the spread across identically prepared states rather than the disturbance caused by a single measurement, which is the picture in Heisenberg's original microscope argument. Experimental tests therefore compare spreads measured on a second sample of particles, which is also prepared in the same state. One neutron-interferometry experiment has been presented as a direct confirmation of the relation between position and momentum.\n\nSeveral foundational questions remain open. The Born rule attaches probabilities to outcomes while the Schrödinger evolution of the wave function is deterministic, and the rule is where definite results appear to emerge from measurement. Attempts to derive the rule from frequency operators, many-worlds branching, Bayesian reasoning, or envariance have been criticized by some authors as circular. Entanglement adds a further puzzle: a composite quantum system can have correlations between its parts that cannot be reproduced by any assignment of definite local values. Violations of Bell inequalities in experiments are generally read as ruling out local hidden-variable explanations, although interpretations differ on what the correlations mean for the nature of probability and reality."
    },
    {
      "heading": "Information Theory and Randomized Computation",
      "text": "Shannon entropy measures the average uncertainty of a discrete random variable \\(X\\) with probability mass function \\(p\\): \\(H(X) = -\\sum_x p(x)\\log_2 p(x)\\), in bits. Claude Shannon introduced the quantity in 1948 as part of the foundation of information theory. A fair coin has entropy 1 bit, a coin with bias 0.9 has about 0.47 bits, and a certain outcome has none. Among distributions on \\(n\\) outcomes, the uniform one has the largest entropy, \\(\\log_2 n\\). Entropy has an operational meaning through the source coding theorem: no lossless code can use fewer bits per symbol on average than \\(H(X)\\), and block coding or Huffman coding can approach that bound. This is the principle behind data compression, in which frequent symbols receive short codewords. For a noisy channel, the mutual information \\(I(X;Y)=H(X)-H(X\\mid Y)\\) measures how much observing the output \\(Y\\) reduces uncertainty about the input \\(X\\). The channel capacity \\(C=\\max_{p(x)} I(X;Y)\\) is the highest rate at which information can be sent with arbitrarily small error, a result known as the noisy-channel coding theorem. For a binary symmetric channel with crossover probability \\(q\\), the capacity is \\(1-H_2(q)\\).\n\nSeveral divergence-type measures compare distributions. The Kullback–Leibler divergence \\(D(p\\|q)=\\sum_x p(x)\\log\\frac{p(x)}{q(x)}\\) is non-negative, vanishes only when \\(p=q\\), and is not symmetric, so it is not a metric. Cross-entropy \\(H(p,q)=-\\sum_x p(x)\\log q(x)\\) decomposes as \\(H(p)+D(p\\|q)\\), which means that minimizing cross-entropy against a fixed \\(p\\) is equivalent to minimizing the divergence from \\(p\\). Mutual information can also be written as the divergence between the joint distribution and the product of the marginals, so it equals zero exactly when the variables are independent. For a bivariate normal pair with correlation \\(\\rho\\), the mutual information is \\(-\\tfrac12\\log(1-\\rho^2)\\), which shows how a linear dependence measure relates to an information-theoretic one.\n\nRandomized algorithms use chance to obtain speed or simplicity that deterministic methods do not easily provide. In randomized quicksort, a pivot chosen uniformly at random yields an expected number of comparisons of about \\(2n\\ln n\\), a result obtained by summing indicator variables over pairs of elements and applying linearity of expected value. Algorithms are commonly classified as Las Vegas algorithms, which always return a correct answer but have random running times, and Monte Carlo algorithms, which run in bounded time and may err with small probability. Examples include the Miller–Rabin primality test, whose error probability for a composite number is at most one quarter per round, and Karger's randomized minimum-cut algorithm, which repeatedly contracts random edges and succeeds with probability at least \\(2/(n(n-1))\\) per run. Expected-time analysis therefore asks for the average cost over the algorithm's internal coin flips, and for a fixed input it is distinct from the average over inputs.\n\nThe Monte Carlo method estimates an expectation \\(\\mathbb{E}[f(X)]\\) by averaging \\(f\\) over independent samples drawn from the target distribution. The law of large numbers guarantees convergence, and the central limit theorem gives the standard error \\(\\sigma/\\sqrt{N}\\) for \\(N\\) samples. This rate does not depend on the dimension of the integral, which is why the method is used for high-dimensional integrals where grid-based quadrature becomes impractical. A simple illustration is the estimation of \\(\\pi\\) by sampling points uniformly in a square and counting the fraction that fall inside the inscribed quarter circle. The method was developed at Los Alamos in the 1940s by Stanisław Ulam and John von Neumann, with the name usually attributed to Nicholas Metropolis, and it is now applied to neutron transport, particle physics, and option pricing.\n\nMarkov chain Monte Carlo (MCMC) addresses the case in which direct sampling is infeasible, typically because a target density \\(\\pi(x)\\) is known only up to a normalizing constant, as with a Boltzmann distribution or a Bayesian posterior. The Metropolis algorithm of 1953 constructs a Markov chain whose stationary distribution is \\(\\pi\\). A proposed move from \\(x\\) to \\(y\\) is accepted with probability \\(\\min\\bigl(1,\\,\\pi(y)q(x\\mid y)/(\\pi(x)q(y\\mid x))\\bigr)\\), where \\(q\\) is the proposal kernel. Because the normalizing constant cancels in this ratio, it never needs to be computed. Hastings generalized the scheme in 1970, and detailed balance together with ergodicity ensures that the chain converges to \\(\\pi\\). Practical difficulties include slow mixing when the distribution has several separated modes, the need to discard an initial burn-in segment, and the use of convergence diagnostics before the samples are trusted. Gibbs sampling is a special case in which each coordinate is updated from its full conditional distribution."
    },
    {
      "heading": "Machine Learning and Artificial Intelligence",
      "text": "Probability enters machine learning mainly through models that return a distribution over possible labels rather than a single answer. Logistic regression passes a weighted sum of input features through a sigmoid function (or a softmax for several classes), so its output can be read as a class probability. Naive Bayes applies Bayes' theorem directly: it multiplies a class prior by per-feature likelihoods, assuming the features are conditionally independent given the class. That assumption is rarely exact, and the resulting numbers are often poorly calibrated even when the ranking of classes is useful.\n\nTraining objectives for classifiers are usually derived from likelihood. For a one-hot label, the cross-entropy sum collapses to one term, the negative logarithm of the probability the model gave to the true class. Since y is one-hot, only one term survives — the negative log-probability of the true class. Under a categorical model this expression is the negative log-likelihood of the observed labels, which links the loss to maximum likelihood estimation and gives it the usual properties of that method, such as consistency and asymptotic efficiency. The binary case corresponds in the same way to a Bernoulli likelihood. The same logic extends beyond discrete outputs: mean squared error in linear regression is the maximum likelihood objective when the target is assumed to be Gaussian, so squared-error and cross-entropy training are two instances of one principle.\n\nStructured dependencies are handled by probabilistic graphical models. A Bayesian network uses a directed acyclic graph in which each variable is conditioned only on its parents, so the joint distribution factorizes into a product of local conditional distributions. A hidden Markov model assumes a latent Markov chain whose states generate the observed sequence; the forward algorithm computes the likelihood of an observation sequence, and posterior decoding infers the most probable hidden states. Gaussian mixture models apply a similar latent-variable structure to clustering and density estimation.\n\nLarge language models assign probabilities to token sequences by the chain rule, \\(P(x_1,\\dots,x_T)=\\prod_{t=1}^{T} P(x_t \\mid x_{<t})\\), so each next-token distribution is a conditional probability. Whether these numbers mean what they appear to mean is a separate question, addressed by calibration. A calibrated model that assigns 0.8 to a set of predictions should see about 80% of them turn out correct. Early studies found that three strong generative models—T5, BART, and GPT-2—and study whether their probabilities on QA tasks are well calibrated, finding the answer is a relatively emphatic no. Miscalibration can also be systematic, since predicted probabilities subject to several forms of biasing, e.g., favoring tokens observed more during pretraining or tokens that appear near the end of a prompt. Post-hoc corrections such as temperature-based scaling are widely used to adjust these outputs, and more recent work reports that Large Language Models (LLMs) have demonstrated inherent calibration capabilities, where predicted probabilities align well with correctness, despite prior findings that deep neural networks are often overconfident.\n\nUncertainty in predictions is addressed in two main ways. Ensemble methods train several models with different initial conditions or resampled training data and combine their outputs; disagreement among members is read as a sign of epistemic uncertainty, and averaging often improves both accuracy and calibration. Bayesian approaches instead place a distribution over model parameters and compute a predictive distribution by averaging over the posterior. Exact averaging is rarely tractable for large networks, so approximations such as variational inference or Markov chain Monte Carlo are used in practice. Calibration itself is often checked with the Brier score, which measures the mean squared difference between predicted probabilities and actual outcomes, so that lower values indicate better calibration."
    },
    {
      "heading": "Applications in Science, Medicine, Genetics, and Social Science",
      "text": "In empirical disciplines, probability does two jobs: it models the chance mechanisms that generate data, and it quantifies how much weight observed differences deserve. Randomization is the clearest case of the first job. In a randomized controlled trial, participants are assigned to treatment or control by a chance device, so each person has a known probability of entering each arm, and allocation does not depend on any personal characteristic. Age, sex, and disease severity therefore tend to be balanced across arms on average, including factors no investigator thought to measure. This balance is what allows an outcome difference to be attributed to the intervention. The balance is only approximate, and small trials can end up with chance imbalances, which is why block randomization and stratification are used. Allocation concealment, which keeps future assignments unpredictable to recruiters, guards against deliberate or unconscious steering of patients into one arm.\n\nProbability also underlies the formal definition of causal effects. In the potential outcomes framework, which goes back to Neyman's work on agricultural experiments in the 1920s and was developed by Donald Rubin in the 1970s, the effect of a treatment on one unit is the difference between the outcome that unit would show under treatment and under control. Only one of these outcomes is ever observed for a given unit, so individual effects cannot be read off directly. Causal inference therefore becomes a missing-data problem, solved by comparing groups under assumptions that randomization secures by design and that observational studies must justify. This is how probability separates association from causation: an observed correlation may arise because a third variable, a confounder, influences both exposure and outcome. The causal inference literature formalizes when such adjustment is legitimate.\n\nIn epidemiology, probabilities are summarized as incidence (the proportion or rate of new cases over a period) and prevalence (the proportion of a population affected at a given time). Comparative measures include the relative risk, the ratio of disease risk in an exposed group to that in an unexposed group, and the odds ratio, which is the standard measure in case-control designs, where the absolute risk cannot be estimated directly. Confidence intervals around these ratios express sampling uncertainty, and the underlying counts are often modeled with the binomial or Poisson distributions.\n\nIn population genetics, the Hardy–Weinberg principle, derived independently by G. H. Hardy and Wilhelm Weinberg in 1908, gives a probabilistic null model. In a large population with random mating and no mutation, selection, migration, or drift, the frequencies of the two genotypes formed from alleles at frequencies p and q settle at p², 2pq, and q², and they stay there from one generation to the next. Departures from these proportions are used to detect evolutionary forces. Genetic drift is the stochastic counterpart: each generation's allele frequencies are a random sample of the gametes that produced them, so frequencies fluctuate by chance. The effect is strongest in small populations, and over long periods it leads to the fixation of one allele and the loss of the other. Under neutrality, the probability that an allele currently at frequency p eventually becomes fixed equals p, a result that illustrates how drift can be analyzed with martingale-type reasoning.\n\nNuclear physics offers a textbook example of a memoryless process. Each radioactive nucleus has a constant probability per unit time of decaying, and this probability does not depend on how long the nucleus has existed. The number of surviving nuclei therefore falls as \\(N(t)=N_0\\,2^{-t/T}\\), where \\(T\\) is the half-life. A single nucleus has a 50% chance of surviving one half-life, whatever its history. Individual decay times cannot be predicted, but for large samples the counts of decays in short intervals are well described by a Poisson distribution, and the waiting times follow an exponential distribution.\n\nOutside the laboratory, the same tools appear in several fields. Demographers and actuaries build life tables from mortality data, in which the probability of dying between two consecutive ages and the probability of surviving to a given age are estimated from observed rates. In reliability engineering, component lifetimes are often modeled with an exponential distribution (constant failure rate) or a Weibull distribution, and system reliability is computed from series and parallel arrangements, usually under an independence assumption. In quality control, Shewhart's control charts, introduced at Bell Labs in the 1920s, use binomial or normal models of a process to set limits beyond which a shift is signaled. Limits placed at three standard deviations from the mean imply a false-alarm rate of roughly 0.27% per point for a stable normal process, a choice that balances missed shifts against spurious alarms."
    },
    {
      "heading": "Economics, Finance, Insurance, and Decision-Making",
      "text": "Expected utility and decision under risk\n\nProbability enters economics mainly through the question of how a rational agent should choose among uncertain prospects. Theory of Games and Economic Behavior, published in 1944 by John von Neumann and Oskar Morgenstern, is considered the groundbreaking text that created the interdisciplinary research field of game theory. Its axiomatic derivation of utility appeared in an appendix to the Second Edition (1947). In the resulting expected utility framework, each lottery is valued as the probability-weighted sum of the utilities of its outcomes, and the agent selects the option with the highest value. Risk aversion then corresponds to a concave utility function: The usual assumption in economics is that decision-makers are risk averse. This means that they have positive but diminishing marginal utility for money, and hence u(x) is increasing and concave.\n\nThe theory was for decades treated as both a normative and a descriptive model, but experiments exposed systematic departures from it. Maurice Allais and Daniel Ellsberg showed that, in certain choice problems, decisions were usually inconsistent with the axioms of expected utility theory. These problems are usually referred to as the Allais paradox and Ellsberg paradox. Leonard Savage's extension allowed the probabilities themselves to be subjective, which is why decision-theory now covers both objective risk and personal belief.\n\nInsurance and pooled risk\n\nThe insurance business rests on the law-of-large-numbers. If an insurer holds \\(n\\) independent policies whose claims each have variance \\(\\sigma^2\\), the average claim per policy has variance \\(\\sigma^2/n\\), so the aggregate result becomes increasingly predictable as the pool grows. A premium is therefore set near the expected claim, with a loading added for administrative costs, capital, and a margin against error. The central-limit-theorem supplies an approximate distribution for total claims, which is used to size reserves and to estimate the probability that losses exceed available capital.\n\nThese results depend on the assumptions behind them. Catastrophic events such as earthquakes or pandemics affect many policyholders at once, which violates independence, and heavy-tailed losses can have an unstable mean or variance. For that reason, actuarial practice combines pooling with reinsurance, catastrophe modelling, and stress testing.\n\nAsset pricing and financial models\n\nContinuous-time models of prices draw on the brownian-motion process. Louis Bachelier's 1900 thesis introduced Brownian motion to model stock prices, but the idea gained lasting influence only later. Samuelson's geometric version, in which the log-price follows a random walk with drift, became the standard model for stock prices. In 1973 Fischer Black and Myron Scholes published a formula for European options. They gave a formula for a call option based on an arbitrage argument and the Itô formula. The key idea is that a portfolio of stock and option can be adjusted continuously so that the risk cancels, which pins down a unique fair price in the Black–Scholes model. Robert Merton developed the underlying mathematics in the same year.\n\nThe model's assumptions are known to be restrictive. It has limitations, including not accounting for heavy tails and excess kurtosis in historical returns. Constant volatility also fails to match observed option prices across strikes, a discrepancy known as the volatility smile, which is why practitioners calibrate richer models to market data.\n\nCredit risk and default probability\n\nProbability of default is a central quantity in lending and bond pricing. In 1974 Robert Merton proposed a structural approach in which a company has a certain amount of zero-coupon debt that will become due at a future time T, and the company defaults if the value of the company's assets falls below the debt; equity is thus a call option on the firm's assets with a strike equal to the face value of the debt. Under the model's assumption that asset values follow a geometric Brownian motion, the default probability at maturity is \\(P(A_T \\le L)\\), which reduces to a normal tail probability \\(N(-d_2)\\) in the Black–Scholes form. This is the logic of the Merton model, and the commercial KMV approach is built on it. The Merton model provides an intuitive way to estimate a firm's default risk by relating its debt and asset value, but its accuracy is limited by simplifying assumptions about how firms and financial markets behave. Estimates obtained this way are often risk-neutral rather than real-world probabilities, so they can differ from observed default frequencies. Reduced-form models, which treat default as the arrival of a random event, are the usual alternative.\n\nGame theory and mixed strategies\n\nIn games with opposing players, probability represents deliberate randomization. A pure strategy can be exploited if the opponent anticipates it, so a player may choose a mixed-strategy, selecting each action with a specified probability. In matching pennies, for example, the unique equilibrium has each player choosing heads or tails with probability one half, which leaves the opponent with no profitable deviation. Nash's 1950 existence result guarantees that every finite game has at least one equilibrium once mixed strategies are allowed. Probability thus serves a dual role here: it describes a player's own randomization and the uncertainty each player has about the other's choice, which is why beliefs about opponents are modelled probabilistically throughout game-theory."
    },
    {
      "heading": "Paradoxes and Conceptual Puzzles",
      "text": "The St. Petersburg paradox\n\nThe paradox was stated by Nicolas Bernoulli in 1713 and published by his cousin Daniel Bernoulli in 1738 in the proceedings of the St Petersburg Academy, which gave it its name. In the game, a fair coin is tossed until it first lands heads. If that happens on toss \\(n\\), the player receives \\(2^{n-1}\\) ducats, an outcome with probability \\(2^{-n}\\). Each possible stage therefore contributes \\(2^{-n}\\cdot 2^{n-1} = \\tfrac12\\) to the expected value, so the sum \\(\\sum_{n\\ge 1}\\tfrac12\\) diverges. A strict expected-value rule would say a player should pay any finite stake to enter, yet few people would pay even a modest sum. Daniel Bernoulli's response was to replace money with utility, assuming that each additional unit of wealth adds less utility than the last, and he proposed a logarithmic form. This makes the expected utility finite. Gabriel Cramer had proposed a related root-based utility. Later work went further: Karl Menger's 1934 argument is often cited for the conclusion that utility must be bounded above. The paradox thus became a central motivation for decision-theory and for the distinction between expected monetary value and expected utility. It also raised the broader question of whether an expected value is a reasonable guide to choice when payoffs are unbounded.\n\nBertrand's paradox\n\nIn 1889 Joseph Bertrand asked for the probability that a randomly chosen chord of a circle is longer than the side of the inscribed equilateral triangle. Three natural procedures give three different answers. If the two endpoints are chosen independently and uniformly on the circle, the probability is 1/3. If the distance of the chord's midpoint from the centre is uniform, the probability is 1/2. If the midpoint is chosen uniformly over the disc, the probability is 1/4. Each procedure implicitly defines a different probability-distribution over the set of chords, so the phrase \"at random\" does not specify a unique model. E. T. Jaynes proposed in 1973 that invariance under symmetry transformations selects the answer 1/2, appealing to a principle of maximum ignorance. Other authors contend that the physical or geometric construction determines the answer, and the debate has not been fully settled. The paradox shows that the principle of indifference is not stable under changes of parametrization, and that a continuous probability problem must specify the measure used.\n\nSimpson's paradox\n\nThe best-known case involves graduate admissions at the University of California, Berkeley in autumn 1973. Across 12,763 applicants, about 44% of men were admitted compared with about 35% of women. When Bickel, Hammel and O'Connell examined the 101 departments separately, they found that women applied disproportionately to departments with low admission rates, and that in most departments the small difference favoured women. The aggregate association thus reversed once the data were split by department. The arithmetic is unproblematic. The interpretation is not: whether the aggregate figure reflects bias depends on the causal role of department choice, which may be a mediator of the effect of sex, a confounder, or something else. Researchers in causal-inference use this case to illustrate why the appropriate level of aggregation must be decided using a causal model rather than the data alone. The same structure appears in medicine, where a treatment can perform better in every subgroup yet worse overall if subgroup membership is linked to treatment assignment, which is why randomization and adjustment for confounding matter.\n\nThe paradox of the great circle\n\nThe Borel–Kolmogorov paradox, which Émile Borel described earlier and Andrey Kolmogorov analysed in 1933, shows the problem of conditioning on an event of probability zero. Suppose a point is uniformly distributed on a sphere, and we ask for its distribution given that it lies on a particular great circle, such as a meridian. The answer depends on which coordinate system is used to describe the circle: conditioning on the longitude of the point, or on a different coordinate that fixes the same circle, yields different conditional densities. Kolmogorov's resolution is measure-theoretic. Conditional probability given a \\(\\sigma\\)-algebra is defined only up to null sets, through the Radon–Nikodym theorem, and it is therefore meaningless to condition on a null event without specifying the \\(\\sigma\\)-algebra or the limiting procedure. The paradox is thus not a flaw in the calculus but a demonstration that the conditioning event must be stated precisely in the probability-space.\n\nThe problem of induction\n\nDavid Hume argued that no amount of observed experience logically entails what will happen in unobserved cases, and that the usual justification relies on an assumption of uniformity that cannot itself be proved from experience without circularity. Probability theory offers a partial reply. Under a uniform prior, Laplace's rule of succession gives a probability of \\((s+1)/(n+2)\\) for success on the next trial after \\(s\\) successes in \\(n\\) trials, so predictions about the unobserved follow from a stated Bayesian model. However, the result depends on the prior, and the choice of prior reintroduces the original question of what justifies assuming a particular pattern. Karl Popper rejected the idea that probability could confirm general hypotheses, and proposed in its place a measure of corroboration. The debate therefore remains open: probabilistic inference can make predictions coherent and quantitative, but it does not by itself settle the question of which background assumptions are warranted."
    },
    {
      "heading": "Open Problems and Current Debates",
      "text": "The first unresolved question is what probability means. The Kolmogorov axioms are compatible with objective readings (long-run frequencies, physical propensities) and with subjective readings (degrees of rational belief), so the formalism does not decide between them. The disagreement concerns single cases and scientific hypotheses. A statement such as \"this drug works with probability 0.7\" has a clear meaning for a repeatable process, but a hypothesis is either true or false, so it has no long-run frequency. Frequentists therefore tie such statements to a reference class, while Bayesians treat them as coherent credences. In practice many researchers combine the two, using objective chances where they exist and subjective judgment elsewhere, and this mixed position has not been formally settled.\n\nThe choice of prior distribution is the central issue in Bayesian analysis. Bayes' theorem gives the posterior as \\(P(H\\mid E)=\\frac{P(E\\mid H)\\,P(H)}{P(E)}\\), so the prior \\(P(H)\\) enters directly. When data are sparse or the model is weakly identified, different reasonable priors can lead to noticeably different conclusions from the same dataset. Responses include sensitivity analysis (reporting how conclusions change across a family of priors), weakly informative priors that constrain implausible values without encoding strong beliefs, and prior predictive checks that test whether a prior generates plausible data. Critics argue that these remedies reduce but do not remove the dependence on modelling choices, and defenders reply that frequentist procedures also embed choices, such as the sampling plan and the test statistic.\n\nReproducibility problems have sharpened the debate over p-values and confidence intervals. In 2016 the American Statistical Association issued its first formal statement on a statistical practice, listing six principles. Among them: a p-value measures how incompatible the data are with a specified model, and it is not the probability that the hypothesis is true or that the results arose by chance alone. The statement also warns against deciding scientific or policy questions on the basis of whether a p-value crosses a threshold. ASA leaders at the time linked the threshold culture to publication bias (the file-drawer effect) and to p-hacking, the search through analyses until a small p-value appears. Proposals since then include abandoning fixed significance thresholds, reporting effect sizes with intervals, using Bayes factors, and preregistering analyses. Related concerns include low statistical power, which inflates the share of significant findings that are false, and uncorrected multiple comparisons. Confidence intervals face a similar misreading: a 95% interval describes the long-run coverage of a procedure, not the probability that one computed interval contains the parameter.\n\nIn quantum mechanics, the Born rule assigns probabilities to measurement outcomes from the wave function, but its status remains contested. Kolmogorov-style \"ignorance\" readings, in which probabilities reflect incomplete knowledge of a pre-existing state, are widely regarded as incompatible with the quantum formalism, and no consensus has emerged on what the rule means. Several research programmes try to derive the rule from more basic assumptions, such as decision-theoretic arguments in the Everett framework or Gleason-type theorems, but critics have argued that some of these derivations are circular or presuppose the rule they aim to explain. Other authors treat the rule as a law whose domain of validity is limited rather than exact. These debates are tied to the measurement problem, which asks how definite outcomes arise at all, and to the interpretation of entanglement correlations.\n\nCalibration is a practical open problem in machine learning. A classifier is calibrated when its stated confidence matches the frequency of correct predictions among cases given that confidence. High accuracy does not guarantee this property. Modern deep networks trained with cross-entropy were found in the late 2010s to be systematically overconfident, and calibration is commonly measured by binned metrics such as expected calibration error. Post-hoc corrections such as temperature scaling are cheap and effective in-distribution, but their benefit shrinks when test data differ from training data, and in extreme shifts they can make matters worse. Some more recent models show underconfidence on in-distribution data, which suggests that the direction of miscalibration depends on training methods and data. Because wrong but confident predictions carry real costs in medical diagnosis and autonomous driving, researchers are developing ensembles, Bayesian approximations, and proper scoring rules to quantify uncertainty. Evaluating the calibration of large language models raises further difficulties, because their outputs are sequences rather than single labels."
    }
  ],
  "links": [
    "large-language-model",
    "statistics",
    "information-theory",
    "statistical-mechanics",
    "quantum-mechanics",
    "pierre-de-fermat",
    "pierre-simon-laplace",
    "andrey-kolmogorov",
    "sample-space",
    "sigma-algebra",
    "measure-theory",
    "odds",
    "statistical-independence",
    "ancient-greece",
    "aristotle",
    "expected-value",
    "law-of-large-numbers",
    "binomial-distribution",
    "normal-distribution",
    "central-limit-theorem",
    "probability-space",
    "random-variable",
    "probability-distribution",
    "bayesian-inference",
    "conditional-probability",
    "bayes-theorem",
    "law-of-total-probability",
    "prior-distribution",
    "posterior-distribution",
    "bayes-factor",
    "thomas-bayes",
    "probability-mass-function",
    "probability-density-function",
    "lebesgue-measure",
    "cumulative-distribution-function",
    "covariance",
    "cauchy-distribution",
    "variance",
    "standard-deviation",
    "correlation",
    "markovs-inequality",
    "chebyshevs-inequality",
    "jensens-inequality",
    "bernoulli-distribution",
    "hypergeometric-distribution",
    "poisson-distribution",
    "exponential-distribution",
    "poisson-process",
    "uniform-distribution",
    "beta-distribution",
    "information-entropy",
    "multivariate-normal-distribution",
    "convergence-in-probability",
    "almost-surely",
    "markov-chain",
    "transition-matrix",
    "stationary-distribution",
    "monte-carlo-method",
    "hidden-markov-model",
    "random-walk",
    "gamblers-ruin",
    "brownian-motion",
    "martingale",
    "conditional-expectation",
    "karl-popper",
    "born-rule",
    "estimator",
    "sampling-distribution",
    "bias-variance-tradeoff",
    "likelihood-function",
    "maximum-likelihood-estimation",
    "ronald-fisher",
    "null-hypothesis",
    "p-value",
    "type-i-and-type-ii-errors",
    "statistical-power",
    "confidence-interval",
    "boltzmann-distribution",
    "partition-function",
    "claude-shannon",
    "mutual-information",
    "kullback-leibler-divergence",
    "cross-entropy",
    "randomized-algorithm",
    "markov-chain-monte-carlo",
    "logistic-regression",
    "naive-bayes-classifier",
    "probabilistic-graphical-model",
    "bayesian-network",
    "gaussian-mixture-model",
    "randomized-controlled-trial",
    "confounding",
    "causal-inference",
    "epidemiology",
    "relative-risk",
    "genetic-drift",
    "half-life",
    "expected-utility-theory",
    "decision-theory",
    "black-scholes-model",
    "mixed-strategy",
    "nash-equilibrium",
    "game-theory",
    "measurement-problem",
    "quantum-entanglement"
  ],
  "sources": [
    {
      "title": "Problem of points",
      "url": "https://en.wikipedia.org/wiki/Problem_of_points"
    },
    {
      "title": "Pat'sBlog: Some additional History on the Problem of Points",
      "url": "https://pballew.blogspot.com/2008/11/some-additional-history-on-problem-of.html"
    },
    {
      "title": "Theory of Probability. A Historical Essay",
      "url": "https://arxiv.org/pdf/1802.09966"
    },
    {
      "title": "The Problem of Points",
      "url": "https://www.usu.edu/math/schneit/StatsHistory/ProbabilityPrompts/Points"
    },
    {
      "title": "Reciprocity as the foundation of Financial Economics",
      "url": "https://arxiv.org/pdf/1310.2798"
    },
    {
      "title": "History of probability",
      "url": "https://en.wikipedia.org/wiki/History_of_probability"
    },
    {
      "title": "The time resolution of the St. Petersburg paradox",
      "url": "https://arxiv.org/pdf/1011.4404"
    },
    {
      "title": "History and Basics of Expected Value Christian Huygens (1629-95)",
      "url": "http://www.sfu.ca/~poitras/Huygens.pdf"
    },
    {
      "title": "CHRISTIANI HUGENII LIBELLUS DE RATIOCINIIS IN LUDO ALEAE OR, The VALUE of all",
      "url": "https://math.dartmouth.edu/~doyle/docs/huygens/huygens.pdf"
    },
    {
      "title": "Expected value",
      "url": "https://en.wikipedia.org/wiki/Expected_value"
    },
    {
      "title": "Chance Combinatorics: The Theory that History Forgot John D. Norton",
      "url": "https://sites.pitt.edu/~jdnorton/papers/chance_combinatorics_final.pdf"
    },
    {
      "title": "1 Review1 of Liber De Ludo Aleae (Book on Games of Chance) by Gerolamo Cardano",
      "url": "https://probability.ca/jeff/ftpdir/mcfadyenreviews.pdf"
    },
    {
      "title": "Ars Conjectandi",
      "url": "https://en.wikipedia.org/wiki/Ars_Conjectandi"
    },
    {
      "title": "Abraham De Moivre",
      "url": "https://www.encyclopedia.com/people/science-and-technology/mathematics-biographies/abraham-de-moivre"
    },
    {
      "title": "Mr. Bayes and the classics: a suggested interpretation",
      "url": "https://arxiv.org/pdf/2202.03819"
    },
    {
      "title": "History of Probability Theory Timeline",
      "url": "https://www.preceden.com/timeline/history-of-probability-theory"
    },
    {
      "title": "Tercentenary of Ars Conjectandi (1713)",
      "url": "https://www.statistics.gov.hk/wsc/IPS008-P1-S.pdf"
    },
    {
      "title": "The Historical Development of Sample Size Estimation: From Huygens and Bernoulli to the Present - The James Lind Library The James Lind Library",
      "url": "https://www.jameslindlibrary.org/articles/the-historical-development-of-sample-size-estimation-from-huygens-and-bernoulli-to-the-present/"
    },
    {
      "title": "The origins and legacy of Kolmogorov's Grundbegriffe",
      "url": "https://arxiv.org/pdf/1802.06071"
    },
    {
      "title": "randomness and foundations of probability: von mises' ...",
      "url": "https://www.ime.usp.br/~walterfm/cursos/mac5796/VanLambalgenOnVonMises.pdf"
    },
    {
      "title": "The Sources of Kolmogorov's Grundbegriffe",
      "url": "https://arxiv.org/pdf/math/0606533"
    },
    {
      "title": "Kolmogorov’s Axiomatisation and its Discontents Aidan Lyon1 January 30, 2012.",
      "url": "https://aidanlyon.com/Bad%20Kolmogrov.pdf"
    },
    {
      "title": "Bayes' theorem",
      "url": "https://en.wikipedia.org/wiki/Bayes'_theorem"
    },
    {
      "title": "Who is the father of Bayesianism: Bayes or Laplace? By Hedibert Lopes1",
      "url": "https://hedibert.org/wp-content/uploads/2026/05/Who-is-the-father-of-Bayesianism.pdf"
    },
    {
      "title": "A nearby galaxy perspective on interstellar dust properties and their evolution",
      "url": "https://arxiv.org/pdf/2202.01868"
    },
    {
      "title": "A Bird's-Eye View of Naming Game Dynamics: From Trait Competition to Bayesian Inference",
      "url": "https://arxiv.org/pdf/2004.01994"
    },
    {
      "title": "Computing Bayes: Bayesian Computation from 1763 to the 21st Century",
      "url": "https://arxiv.org/pdf/2004.06425"
    },
    {
      "title": "On p-Values and Statistical Significance",
      "url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9917591/"
    },
    {
      "title": "AMERICAN STATISTICAL ASSOCIATION RELEASES STATEMENT ON",
      "url": "https://www.amstat.org/asa/files/pdfs/p-valuestatement.pdf"
    },
    {
      "title": "Full article: The ASA Statement on p-Values: Context, Process, and Purpose",
      "url": "https://www.tandfonline.com/doi/full/10.1080/00031305.2016.1154108"
    },
    {
      "title": "ASA President’s Task Force Statement on Statistical Significance and Replicability",
      "url": "https://hdsr.mitpress.mit.edu/pub/50vl2b07"
    },
    {
      "title": "Boltzmann distribution",
      "url": "https://en.wikipedia.org/wiki/Boltzmann_distribution"
    },
    {
      "title": "Boltzmann Distribution -- from Wolfram MathWorld",
      "url": "https://mathworld.wolfram.com/BoltzmannDistribution.html"
    },
    {
      "title": "Born rule — Grokipedia",
      "url": "https://grokipedia.com/page/Born_rule"
    },
    {
      "title": "Mysterious Quantum Rule Reconstructed From Scratch",
      "url": "https://www.quantamagazine.org/the-born-rule-has-been-derived-from-simple-physical-principles-20190213/"
    },
    {
      "title": "Entangled in Spacetime",
      "url": "https://arxiv.org/pdf/2409.02354"
    },
    {
      "title": "A closer look at the uncertainty relation of position and momentum",
      "url": "https://arxiv.org/pdf/0811.2582"
    },
    {
      "title": "Born rule",
      "url": "https://en.wikipedia.org/wiki/Born_rule"
    },
    {
      "title": "Why Probability in Quantum Mechanics is Given by the Wave Function Squared",
      "url": "https://preposterousuniverse.com/blog/2014/07/24/why-probability-in-quantum-mechanics-is-given-by-the-wave-function-squared/comment-page-2/"
    },
    {
      "title": "(PDF) A simple proof of Born's rule for statistical interpretation of quantum mechanics",
      "url": "https://www.academia.edu/31028272/A_simple_proof_of_Borns_rule_for_statistical_interpretation_of_quantum_mechanics"
    },
    {
      "title": "Heisenberg's original derivation of the uncertainty principle and its universally valid reformulations",
      "url": "https://arxiv.org/pdf/1507.02010"
    },
    {
      "title": "Cross-Entropy Loss for Classification",
      "url": "https://www.abhik.ai/concepts/deep-learning/cross-entropy-loss"
    },
    {
      "title": "How Can We Know When Language Models Know? On the Calibration of Language Models for Question Answering",
      "url": "https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00407/107277/How-Can-We-Know-When-Language-Models-Know-On-the"
    },
    {
      "title": "Language Models in the Loop: Incorporating Prompting into Weak Supervision",
      "url": "https://arxiv.org/pdf/2205.02318"
    },
    {
      "title": "Calibration Across Layers: Understanding Calibration Evolution in LLMs",
      "url": "https://arxiv.org/html/2511.00280v1"
    },
    {
      "title": "Model Calibration, Explained: A Visual Guide with Code Examples for Beginners",
      "url": "https://towardsdatascience.com/model-calibration-explained-a-visual-guide-with-code-examples-for-beginners-55f368bafe72/"
    },
    {
      "title": "Generative Cross-Entropy: A Strictly Proper Loss for Data-Efficient Classification",
      "url": "https://arxiv.org/pdf/2604.06689"
    },
    {
      "title": "Uses and Abuses of the Cross-Entropy Loss: Case Studies in Modern Deep Learning",
      "url": "https://arxiv.org/pdf/2011.05231"
    },
    {
      "title": "A Gentle Introduction to Cross-Entropy for Machine Learning - MachineLearningMastery.com",
      "url": "https://machinelearningmastery.com/cross-entropy-for-machine-learning/"
    },
    {
      "title": "Randomized controlled trials",
      "url": "https://www.sciencedirect.com/science/article/pii/S147751312400617X"
    },
    {
      "title": "Randomized controlled trials",
      "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC4910682/"
    },
    {
      "title": "Randomized Control Trial (RCT) - Simply Psychology",
      "url": "https://www.simplypsychology.org/randomized-controlled-trial.html"
    },
    {
      "title": "Hardy-Weinberg equilibrium",
      "url": "https://www.nature.com/scitable/definition/hardy-weinberg-equilibrium-122/"
    },
    {
      "title": "Rethinking (again) Hardy-Weinberg and genetic drift in undergraduate biology - PMC",
      "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC10285527/"
    },
    {
      "title": "Hardy-Weinberg and Genetic Drift in Biology - 31K+ Views",
      "url": "https://www.jove.com/science-education/v/10559/hardy-weinberg-equilibrium-principle-and-genetic-drift-concept"
    },
    {
      "title": "Law of radioactive decay",
      "url": "https://www.walter-fendt.de/html5/phen/lawdecay_en.htm"
    },
    {
      "title": "Astrophysics with radioactive isotopes",
      "url": "https://arxiv.org/pdf/1811.10604"
    },
    {
      "title": "Half-Life and Activity",
      "url": "https://courses.lumenlearning.com/suny-physics/chapter/31-5-half-life-and-activity/"
    },
    {
      "title": "Theory of Games and Economic Behavior",
      "url": "https://en.wikipedia.org/wiki/Theory_of_Games_and_Economic_Behavior"
    },
    {
      "title": "Mean-Variance and Expected Utility: The Borch Paradox",
      "url": "https://arxiv.org/pdf/1306.2728"
    },
    {
      "title": "Generalized expected utility",
      "url": "https://en.wikipedia.org/wiki/Generalized_expected_utility"
    },
    {
      "title": "The Black-Scholes Option Pricing Model A Comprehensive Analysis Renoir Vieira",
      "url": "https://portal.if.usp.br/grenac/sites/portal.if.usp.br.ifusp/files/black_scholes_presentation.pdf"
    },
    {
      "title": "The Golden Age of the Mathematical Finance",
      "url": "https://arxiv.org/pdf/2102.06693"
    },
    {
      "title": "Black-Scholes Model - an overview",
      "url": "https://www.sciencedirect.com/topics/mathematics/black-scholes-model"
    },
    {
      "title": "(PDF) Merton's Model, Credit Risk and Volatility Skews",
      "url": "https://www.researchgate.net/publication/228294160_Merton's_Model_Credit_Risk_and_Volatility_Skews"
    },
    {
      "title": "Merton model",
      "url": "https://en.wikipedia.org/wiki/Merton_model"
    },
    {
      "title": "Expected Utility: Von Neumann-Morgenstern, Risk Aversion, Independence Axiom",
      "url": "https://theorempath.com/topics/expected-utility"
    },
    {
      "title": "Risk, ambiguity and quantum decision theory",
      "url": "https://arxiv.org/pdf/0711.0886"
    },
    {
      "title": "Default Probability by Using the Merton Model for Structural Credit Risk - MATLAB & Simulink",
      "url": "https://www.mathworks.com/help/risk/default-probability-using-the-merton-model-for-structural-credit-risk.html"
    },
    {
      "title": "Default prediction with the Merton-type structural model based on the NIG Lévy process - ScienceDirect",
      "url": "https://www.sciencedirect.com/science/article/pii/S0377042716303673"
    },
    {
      "title": "The St. Petersburg Paradox (Stanford Encyclopedia of Philosophy)",
      "url": "https://plato.stanford.edu/entries/paradox-stpetersburg/"
    },
    {
      "title": "Relative Net Utility and the Saint Petersburg Paradox",
      "url": "https://arxiv.org/pdf/1910.09544"
    },
    {
      "title": "The Saint Petersburg Paradox and Its Solution",
      "url": "https://www.mdpi.com/2227-9091/13/2/32"
    },
    {
      "title": "The St. Petersburg Paradox (Stanford Encyclopedia of Philosophy/Spring 2005 Edition)",
      "url": "https://plato.stanford.edu/archIves/spr2005/entries/paradox-stpetersburg/"
    },
    {
      "title": "Experimental tests of Bertrand's question and the Duhem-Quine problem",
      "url": "https://arxiv.org/pdf/1806.00308"
    },
    {
      "title": "Bertrand's Paradox",
      "url": "https://www.omnicalculator.com/statistics/betrand-paradox"
    },
    {
      "title": "Bertrand paradox (probability)",
      "url": "https://en.wikipedia.org/wiki/Bertrand_paradox_(probability)"
    },
    {
      "title": "Simpson's paradox",
      "url": "https://en.wikipedia.org/wiki/Simpson's_paradox"
    },
    {
      "title": "Revisiting the Berkeley Admissions data: Statistical Tests for Causal Hypotheses",
      "url": "https://arxiv.org/pdf/2502.10161"
    },
    {
      "title": "Statistical significance and p-values",
      "url": "https://www.sciencedaily.com/releases/2016/03/160307092305.htm"
    },
    {
      "title": "Quantum Probability: a reliable tool for an agent or a reliable source of reality?",
      "url": "https://arxiv.org/pdf/1903.03863"
    },
    {
      "title": "The Born rule",
      "url": "https://arxiv.org/pdf/2502.08545"
    },
    {
      "title": "Updating the Born rule",
      "url": "https://arxiv.org/pdf/1702.01845"
    },
    {
      "title": "Foundations of quantum physics I. A critique of the tradition",
      "url": "https://arxiv.org/pdf/1902.10778"
    },
    {
      "title": "Quantifying Calibration Error in Modern Neural Networks through Evidence Based Theory",
      "url": "https://arxiv.org/html/2411.00265v1"
    },
    {
      "title": "Beyond Overconfidence: Model Advances and Domain Shifts Redefine Calibration in Neural Networks",
      "url": "https://arxiv.org/html/2506.09593v1"
    },
    {
      "title": "Rethinking Calibration of Deep Neural Networks: Do Not ...",
      "url": "https://proceedings.neurips.cc/paper/2021/file/61f3a6dbc9120ea78ef75544826c814e-Paper.pdf"
    }
  ],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}