# Statistics Statistics is the discipline of learning from data in the presence of [[Uncertainty|uncertainty]] — and, just as centrally, of saying how wrong you might be. It runs [[Probability_theory|probability theory]] in reverse: probability deduces what data a known [[Probability_distribution|distribution]] would produce, while statistical inference induces which distributions could plausibly have produced the data at hand. That inversion powers [[Estimation_theory|estimation]], hypothesis testing, and experimental design, and makes the field the working epistemology of [[Science|science]] and engineering alike: [[Reliability_engineering|reliability]] life tests, [[Control_theory|control]] loops, [[Econometrics|econometric]] models, clinical trials, and [[Machine_learning|machine learning]] are all, underneath, claims about distributions fitted from finite samples. The core discovery is quantitative humility — averages of n independent measurements fluctuate like 1/√n, so certainty is purchasable, but the price schedule is fixed. ## Population, sample, and the √n engine The founding move is to distinguish the population (everything you care about) from the sample (what you actually measured) and to treat the sample as a draw from a [[Probability_distribution|probability distribution]]. Two limit theorems make this useful. Jacob Bernoulli's law of large numbers (published 1713) guarantees that sample frequencies converge to true probabilities; the central limit theorem — glimpsed by de Moivre in 1733, generalized by Laplace around 1810 — says the error of an average is itself approximately Gaussian with standard deviation σ/√n. This is the √n engine of all empirical [[Science|science]]: quadrupling the data halves the error bar. It is why a well-designed poll of about 1,000 respondents carries a margin near ±3 percentage points regardless of national population, why [[Monte_Carlo_method|Monte Carlo]] estimates converge at rate n^(−1/2), and why [[Signal-to-noise_ratio|signal-to-noise ratio]] climbs with averaging in any [[Sensor|sensor]] system. The [[Expected_value|expected value]] is the target; the sampling distribution is the honest report of scatter around it. ## Estimation: least squares and likelihood Estimation theory began with astronomy. Legendre published least squares in 1805 and Gauss claimed prior use in his 1809 orbit work, giving [[Applied_mathematics|applied mathematics]] its first general recipe for fusing noisy observations: choose parameters minimizing summed squared residuals. R. A. Fisher's 1922 foundations paper delivered the modern vocabulary — maximum likelihood, sufficiency, efficiency — and tied the best achievable precision of any unbiased estimator to the curvature of the log-likelihood, a quantity now called Fisher information with deep formal links to [[Information_theory|information theory]]. Regression descends from Galton's 1880s studies of heredity, where regression to the mean was first isolated; Sewall Wright's [[Path_analysis_(statistics)|path analysis]] (circa 1920) pushed regression toward causal diagrams decades before modern graphical models. The recursive version of least squares matters most to this hub: the [[Kalman_filter|Kalman filter]] (1960) is sequential [[Estimation_theory|estimation]] under [[Feedback|feedback]], weighting each new measurement by inverse variance — the mathematical heart of navigation, [[Control_theory|control]], and [[Time_series|time-series]] tracking. ## Testing, randomization, and the design of experiments Fisher's Rothamsted work in the 1920s and his 1935 *Design of Experiments* established that randomization is what licenses causal talk: assign treatments by chance and confounders average out, turning agriculture, medicine, and now every A/B-tested product into experimental sciences. Neyman and Pearson (1933) formalized testing as a [[Decision_theory|decision problem]] with two error rates — false alarm α and miss β — a framing any [[Radar|radar]] or [[Fault_tree_analysis|fault-detection]] engineer will recognize as their own. Industrial practice added Walter Shewhart's control chart (1924, at the Bell System — the ancestor culture of [[Bell_Labs]]), which watches a [[Manufacturing|manufacturing]] process and signals only when variation exceeds what chance predicts; statistical process control remains the backbone of [[Quality_assurance|quality assurance]] and modern [[Industrial_engineering|industrial engineering]]. The recurring systems lesson: a test is a [[Decision-making|decision rule]] wired into a [[Feedback|feedback]] loop, and its error rates are design parameters, not afterthoughts. ## The Bayesian turn Bayes' theorem (published posthumously, 1763; wielded systematically by Laplace) treats parameters themselves as uncertain quantities: posterior ∝ prior × likelihood. For most of the twentieth century the approach was sidelined by computational cost and by frequentist scruples about priors. Two things revived it: hierarchical models that pool strength across related units, and Markov chain [[Monte_Carlo_method|Monte Carlo]] — descended from the 1953 Metropolis algorithm of [[Statistical_mechanics|statistical-mechanics]] provenance — which made posteriors computable for realistic models. [[Bayesian_network|Bayesian networks]] factor joint distributions over [[Directed_graph|directed graphs]] and carried probabilistic reasoning into [[Artificial_intelligence|artificial intelligence]]; spam filters, medical diagnosis, and sensor fusion in [[Robotics|robotics]] are posterior updates in production. The frequentist–Bayesian argument has settled into engineering pragmatism: use the machinery whose guarantees match the [[Decision_theory|decision]] at hand. ## Statistics inside engineered systems This hub's portals mark where the discipline is structural rather than decorative. [[Reliability_engineering|Reliability engineering]] estimates lifetime distributions from censored test data — the Weibull family (1951) dominates because it models weakest-link failure — and feeds the basic-event probabilities that [[Fault_tree_analysis|fault tree analysis]] and [[Failure_mode_and_effects_analysis|FMEA]] combine into system risk. [[Control_theory|Control]] and [[Signal_processing|signal processing]] rest on statistical models of noise, from the [[Wiener_filter|Wiener filter]] (1940s) to the [[Kalman_filter|Kalman filter]] in every aircraft. [[Queueing_theory|Queueing]] and [[Operations_research|operations research]] models are fitted, not assumed. [[Econometrics]] disciplines economic claims; [[Analytics|analytics]] platforms and the [[List_of_data_science_software|data-science toolchain]] industrialize the workflow; [[Machine_learning|machine learning]] is, on its theoretical side, high-dimensional estimation with prediction loss in place of parameters. Everywhere the pattern repeats: measurement, model, residual, [[Feedback|correction]]. ## Honest caveats The field's power comes bundled with standing failure modes. Correlation is not causation, and Simpson's paradox (named from a 1951 paper) shows aggregate associations reversing inside every subgroup — the modifiable-lens cousin of problems any [[Systems_theory|systems]] thinker should expect. Selection effects bite hardest when invisible: Abraham Wald's wartime survivability analysis — armor where returning bombers show no holes — is the canonical parable. Multiple testing and flexible analysis inflate false positives, a mechanism behind the replication troubles documented across psychology and biomedicine in the 2010s; preregistration, effect-size reporting, and explicit [[Uncertainty|uncertainty]] statements are the countermeasures. The mature stance treats every published number as an [[Estimation_theory|estimate]] with a distribution attached — statistics not as a stamp of certainty, but as [[Decision-making|decision-grade]] bookkeeping of doubt. **On the spine:** [[Probability_theory]] · [[Probability_distribution]] · [[Estimation_theory]] · [[Decision_theory]] · [[Uncertainty]]. ## Wikipedia : Wikitube **Strict pair:** [Wikipedia](https://en.wikipedia.org/wiki/Statistics) : [Wikitube](https://en.wikitube.io/wiki/Statistics) ## Previous hub tags Hubs: `Systems`. Portals: [[PORTAL_Systems]], [[PORTAL_Reliability_engineering]], [[PORTAL_Control_theory]], [[PORTAL_Information_theory]], [[PORTAL_Fault_tree_analysis]]. --- *Repopulated 2026-08-12 · redlink fill · 0 deletions.*