Bayes' Theorem: The Equation That Came Back from the Dead
Thomas Bayes never published his most famous result. A friend found it among his papers after he died.
Thomas Bayes was an English minister and amateur mathematician who died in 1761 without publishing the theorem that bears his name. His friend Richard Price found an unpublished essay in Bayes' papers, recognised its significance, revised it substantially, and submitted it to the Royal Society in 1763. The paper was titled "An Essay towards Solving a Problem in the Doctrine of Chances." Bayes himself may never have considered it worth publishing.
The problem Bayes was solving was inversion: given that you have observed certain evidence, what can you infer about the probability of the cause? His core formula says that the probability of a hypothesis given evidence equals the prior probability of that hypothesis multiplied by the probability of seeing the evidence if the hypothesis were true, divided by the total probability of the evidence. P(H|E) = P(E|H) × P(H) / P(E). This is how you update beliefs in light of new information.
The theorem was largely ignored for most of the 18th and early 19th centuries. Pierre-Simon Laplace independently rediscovered and extended it, putting it on rigorous mathematical footing and applying it to everything from celestial mechanics to mortality statistics. What we now call "Bayesian" statistics is largely Laplace's reconstruction. Bayes wrote the seed; Laplace grew the tree.
For most of the 20th century, Bayesian statistics was considered philosophically suspect by the dominant frequentist school. Frequentists argued that probability should only describe long-run frequencies, not degrees of belief. The "prior," Bayes' starting assumption before seeing evidence, was dismissed as unscientific subjectivity. The Bayesian-frequentist debate was one of the bitterest in the history of science, and it is still not entirely resolved.
Today Bayes' theorem is embedded in spam filters (updating the probability that an email is spam given its words), medical diagnosis (updating the probability of disease given test results), speech recognition, Bayesian neural networks, and the posterior sampling methods at the heart of modern MCMC. The minister who did not publish his own essay is now cited whenever a model updates its beliefs in light of new evidence, which in 2025, is several billion times per second.
What it actually meant
Bayes' theorem codified something humans do intuitively but machines do unreliably: update a belief in proportion to evidence. The prior is not a weakness of the framework: it is honest accounting of what you knew before looking at the data. The debate is not Bayesian versus frequentist. It is about whether uncertainty should be expressed as a parameter or a distribution.
What enterprises get wrong today
Enterprises treat model outputs as probabilities when models are typically miscalibrated. A classifier outputting 0.87 is not asserting 87% probability that the instance belongs to that class: it is outputting a score that may not correspond to empirical frequency. Teams deploy models without calibration checks and use uncalibrated scores to make resource allocation decisions. The downstream decisions are only as good as the calibration.
Related capabilities
More from Mathematics
Ask about this story
Bayes never published his own theorem. His friend found it after he died, polished it, and sent it to the Royal Society. Two hundred and sixty years later it's the mathematical foundation of spam filters, medical diagnosis, and modern AI. What would you like to explore?