# Probability distribution A probability distribution is the complete specification of a random quantity: which values it can take and with what probabilities — everything [[Probability_theory|probability theory]] knows about a variable, packed into one object. Discrete variables carry a probability mass function, continuous ones a [[Probability_density_function|density]], and both are subsumed by the cumulative distribution function F(x) = P(X ≤ x), from which flow the [[Expected_value|expected value]], the variance, and every quantile a [[Statistics|statistician]] or [[Reliability_engineering|reliability engineer]] will ever report. Distributions are where mechanism meets measurement: coin-counting produces the binomial, memoryless waiting produces the exponential, aggregation produces the Gaussian, and [[Statistical_mechanics|statistical mechanics]] writes whole theories of matter as distributions over [[Microstate_(statistical_mechanics)|microstates]]. Reading a system's distribution — especially its tails — is this hub's fastest diagnostic of what kind of [[Complex_system|system]] you are holding. ## One function carries everything The cumulative function F is the distribution's legal identity: monotone, right-continuous, running from 0 to 1. Its derivative (when it exists) is the [[Probability_density_function|density]] f; its generalized inverse F⁻¹ is the quantile function that turns percentiles into values; its integrals against test functions yield moments — the mean as first moment, variance as centered second, with skewness and kurtosis grading asymmetry and tail weight. Summary numbers compress honestly only when the shape cooperates: a mean without a shape claim is an invitation to error, since wildly different laws share means. The empirical distribution of a sample converges to the true F as data accumulate, which is the licensing theorem behind treating histograms as [[Estimation_theory|estimates]] and behind every [[Monte_Carlo_method|Monte Carlo]] summary. In multiple dimensions, joint distributions add dependence structure, and [[Bayesian_network|Bayesian networks]] factor them over [[Directed_graph|directed graphs]] to keep the bookkeeping tractable. ## Mechanism generates family The named families are not a zoo but a set of theorems about generating mechanisms. Count successes in n independent trials: binomial. Count rare events in a window: Poisson (1837), P(k) = λᵏe^(−λ)/k!, which Bortkiewicz famously fitted in 1898 to Prussian cavalry deaths by horse kick. Wait for the first event of a memoryless process: exponential, the only continuous law with P(T > s+t | T > s) = P(T > t) — which is why [[Radioactive_decay|radioactive decay]] is exponential and a [[Half-life|half-life]] t½ = ln 2/λ is a complete description. Add many small independent effects: normal, by the central limit theorem (de Moivre 1733, Gauss 1809). Multiply many small effects: lognormal, the default for incomes, city sizes, and particle grindings. Take the weakest of many links: Weibull (1951), the workhorse lifetime law of [[Reliability_engineering|reliability engineering]] and [[Fault_tree_analysis|failure analysis]]. Queue arrivals at random: Erlang's telephone-era laws, still running [[Queueing_theory|queueing theory]]. | Family | Support | Mechanism | Canonical habitat | | --- | --- | --- | --- | | Binomial | 0…n | independent trials | [[Quality_assurance|acceptance sampling]] | | Poisson | 0, 1, 2, … | rare events per window | arrivals, defects, mutations | | Exponential | t ≥ 0 | memoryless waiting | [[Radioactive_decay|decay]], failure times | | Normal | ℝ | many small additive causes | measurement error, [[Signal-to-noise_ratio|noise]] | | Lognormal | x > 0 | many multiplicative causes | incomes, file sizes | | Weibull | t ≥ 0 | weakest link of many | [[Reliability_engineering|life testing]] | | Power law | x ≥ x_min | preferential growth | [[Scale-free_network|network degrees]], quakes | | Boltzmann | energy states | thermal equilibrium | [[Statistical_mechanics|matter]] | ## Physics runs on distributions Kinetic theory's founding act was distributional: [[James_Clerk_Maxwell|Maxwell]] derived the molecular speed law in 1860, and [[Ludwig_Boltzmann|Boltzmann]] generalized it to the exponential weighting of states by energy, p ∝ e^(−E/k_BT), with k_B = 1.380649 × 10⁻²³ J/K now exact by definition. [[Josiah_Willard_Gibbs|Gibbs's]] ensembles made the distribution-over-[[Microstate_(statistical_mechanics)|microstates]] the very definition of a thermodynamic phase, with [[Entropy|entropy]] measuring its breadth. Quantum statistics then split matter in two by occupancy law: Bose–Einstein (1924–25), whose pile-up into a single ground state becomes a [[Bose–Einstein_condensate|condensate]] near absolute zero, versus Fermi–Dirac (1926), whose exclusion principle props up white dwarfs and [[Semiconductor_device|semiconductor]] band filling. [[Phase_transition|Phase transitions]] appear as non-analytic changes in these distributions at [[Critical_point_(thermodynamics)|critical points]] — macroscopic drama from a shift in probabilistic weight, the purest systems moral in [[Physics|physics]]. ## Tails decide everything Distributions with the same middle can price catastrophe a millionfold apart, so the tail is the risk-relevant organ. Light-tailed laws (normal, exponential) make large deviations essentially impossible; heavy-tailed laws make them routine. Pareto's 1890s income law and its power-law kin p(x) ∝ x^(−α) govern earthquake energies, war sizes, and the degree counts of [[Scale-free_network|scale-free networks]] — the Barabási–Albert growth model of 1999 yields exponent 3 exactly — and under such laws the sample mean converges slowly or, as with the Cauchy distribution, not at all. Practical consequences cascade: [[Statistics|averages]] mislead, variance may be infinite, worst cases dominate design, and [[Uncertainty|uncertainty]] statements built on Gaussian habits fail precisely when they matter. [[Reliability_engineering|Reliability]] practice therefore fits tails deliberately (extreme-value theory descends from Fisher–Tippett 1928), and [[Network_science|network science]] reads a degree distribution's tail as a vulnerability report — heavy tails mean hubs, and hubs mean targeted failure modes for the [[Electrical_grid|grid]] and contagion alike. ## Honest ignorance and manufactured randomness Which distribution should you assume when you know only a little? Jaynes's 1957 maximum-entropy principle answers: the one maximizing [[Entropy_(information_theory)|Shannon entropy]] subject to what you actually know — uniform given only a finite range, exponential given a positive mean, Gaussian given mean and variance. That is why these families are ubiquitous: they are the least-committed laws consistent with common constraints, a bridge [[Claude_Shannon|Shannon's]] [[Information_theory|information theory]] built between inference and [[Statistical_mechanics|thermodynamics]]. The reverse direction is manufacturing: given a target law, draw uniform U and apply F⁻¹(U), or use the 1958 Box–Muller trick for Gaussians, and a [[Simulation|simulation]] can propagate whole distributions through models no formula penetrates — the daily bread of [[Monte_Carlo_method|Monte Carlo]] engineering, [[Machine_learning|machine-learning]] generative models, and [[Decision_theory|decision]] analyses that report risk as a distribution rather than a number, which is what mature [[Uncertainty|uncertainty]] handling looks like. **On the spine:** [[Probability_theory]] · [[Probability_density_function]] · [[Statistics]] · [[Statistical_mechanics]] · [[Monte_Carlo_method]]. ## Wikipedia : Wikitube **Strict pair:** [Wikipedia](https://en.wikipedia.org/wiki/Probability_distribution) : [Wikitube](https://en.wikitube.io/wiki/Probability_distribution) ## Previous hub tags Hubs: `Systems`. Portals: [[PORTAL_Systems]], [[PORTAL_Monte_Carlo_method]], [[PORTAL_Graph_theory]], [[PORTAL_Decision_theory]], [[PORTAL_Reliability_engineering]]. --- *Repopulated 2026-08-12 · redlink fill · 0 deletions.*