Bayesian probability

Core components

Bayesian reasoning centers on four elements: prior, likelihood, evidence and posterior. The prior expresses what is believed before seeing the current data. The likelihood describes how probable the observed data are under different hypotheses. The evidence (or marginal likelihood) normalizes possibilities across all hypotheses. The posterior combines prior and likelihood to give the updated belief after observing data, typically computed via Bayes' theorem: P(H|D) = P(D|H) P(H) / P(D).

How it works in practice

Practically, Bayesian updating is iterative: each posterior can serve as the prior for the next update when new data arrive. This makes the framework natural for sequential learning and decision making. For an intuitive example, consider medical testing: a test result modifies the clinician's initial estimate (prior) of disease probability into a more informed probability (posterior) that accounts for test accuracy (likelihood). For a guided primer, see Bayesian updating.

History and development

The approach takes its name from Thomas Bayes, an 18th-century statistician whose work was expanded by Pierre-Simon Laplace. It has evolved substantially since, integrating with modern probability theory, computation and decision theory. The rise of computing power and numerical methods such as Markov chain Monte Carlo made complex Bayesian models practical for many applications.

Applications and importance

• Machine learning: probabilistic modelling, Bayesian networks, and uncertainty quantification.
• Science and engineering: parameter estimation, model comparison, and hierarchical models.
• Medicine and diagnostics: interpreting test results and combining clinical data.
• Everyday decision making: updating beliefs in light of new information and evidence.

Distinctions and notable points

Bayesian methods differ from frequentist approaches, which define probability in terms of long-run frequencies of repeatable experiments. Critics sometimes object to the subjective element in choosing priors; proponents note that priors can be made explicit and tested, and objective or weakly informative priors are often used. Both paradigms offer valuable tools, and in many practical problems practitioners combine ideas from each. For a comparison with the frequentist viewpoint, see frequentist probability.