The History of AI

Maximum likelihood

Era 1 · The foundations (1763–1942) › Probability and statistics

Lineage
Data and learning
Kind
Theory

Who

What happened

Ronald Fisher publishes "On the mathematical foundations of theoretical statistics" in the Philosophical Transactions of the Royal Society; it had been read before the Society in November 1921.

Why it mattered

Maximum likelihood and the foundations of inferential statistics: the concepts of parameter, statistic, consistency, efficiency, and sufficiency. Training a model on data is still, very often, maximizing its likelihood.

The full story

At the start of the twentieth century, statistics was a collection of techniques with no common theory. The English mathematician Ronald Fisher set out to give it one. The 1922 paper, some sixty pages long, was submitted in June 1921, read before the Royal Society on November 17, and published on April 19, 1922.

Fisher divides the statistician's work into three problems. Specification chooses the family of distributions that describes the data. Estimation calculates, from a sample, the values of that distribution's unknown quantities. Distribution studies how much those calculations vary from one sample to another. For precision, the paper introduces two terms still in use: parameter, the unknown quantity of the population, and statistic, what is calculated from the sample.

Fisher's method of estimation is maximum likelihood: choose the parameters under which the observed data are most probable. Gauss, Laplace, and Edgeworth had used it, but between 1912 and 1922 Fisher turned it into a general method and defined the properties a good estimator should have: consistency, efficiency, and sufficiency. In 1938 Samuel Wilks published the proof now known as Wilks's theorem, which gave the method a justification beyond intuition.

The paper, and the later work of Fisher and Neyman, pushed Laplace's methods of inverse probability aside for a quarter of a century. Today, training a machine learning model often comes down to maximizing its likelihood.

Under the hood

The likelihood of a parameter value θ is the probability of the observed data if θ were the true value: L(θ) = P(data | θ). The maximum likelihood estimator is the θ that makes it largest. For convenience one maximizes its logarithm, which turns the product of the probabilities of each data point into a sum.

If the errors follow a normal distribution, maximizing the likelihood is the same as minimizing the sum of squared errors: least squares is a special case. Training a language model by minimizing cross-entropy is also maximizing the likelihood of the training text.

Builds on

  • Théorie analytique des probabilités (1812)

    Inherited, with no proven direct link

    Laplace and Gauss had used maximum likelihood in particular cases; Fisher turns it into a general method of estimation

Sources

How to cite

“Maximum likelihood.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/fisher-maximum-likelihood

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline