conjugate prior

A conjugate prior is a lucky mathematical match: a prior whose shape, after you fold in the data, comes back out as the same shape — just with updated numbers. It is like having a savings account where every deposit keeps the balance in the same currency. You never have to convert; updating is as simple as adjusting a couple of counters.

Concretely, if your prior belief about a coin's bias is a Beta distribution and you observe coin flips (a Binomial process), the posterior is again a Beta distribution. To update, you just add the number of heads to one parameter and the number of tails to the other. No integrals, no sampling — the Bayesian update collapses into arithmetic. Famous pairs include Beta with Binomial, Gamma with Poisson, and a Gaussian prior with a Gaussian likelihood.

Why it matters, and where it stops: in the era before fast computers, conjugacy was often the only way to do Bayesian inference at all, and it is still prized for being exact, instant, and easy to reason about. But it is a convenience, not a law of nature. Real problems rarely hand you a conjugate model, and forcing one can distort your honest beliefs. When no conjugate prior fits, you fall back on approximate methods like MCMC or variational inference. Treat conjugacy as a gift when it appears, not a requirement to chase.