maximum a posteriori (MAP) estimation
/ MAK-suh-mum ah pos-TEER-ee-OR-eye /
Maximum a posteriori estimation, or MAP, is the move you make when you have a full Bayesian posterior — a whole landscape of plausible answers — but you just want one number to act on. MAP picks the single peak: the most probable value once both your prior and your data have had their say. It is the highest point on the posterior's hill.
Contrast it with its cousin, maximum likelihood, which picks the value that best fits the data alone. MAP adds the prior into the mix, which acts as a gentle pull toward what you believed beforehand. That pull is exactly what stops a model from over-committing to a small or noisy dataset. In fact, many everyday techniques are secretly MAP in disguise: the 'regularization' penalties that keep machine-learning models from overfitting correspond precisely to particular priors.
The honest catch is that MAP throws away most of what makes Bayesian thinking valuable. By reporting only the peak, it discards the width of the distribution — the very thing that told you how uncertain you were. Worse, the peak can be a misleading summary: a posterior can be lopsided or have a tall, thin spike beside a broad, fat region where most of the real probability actually lives. MAP is a useful, cheap point estimate, but if you care about uncertainty, do not mistake the peak for the whole story.
Estimating a basketball player's true free-throw rate after she makes 1 of her first 2 shots. Maximum likelihood says 1/2 = 50%, ignoring everything you know. MAP, with a prior that pros usually shoot around 75%, pulls the estimate up to maybe 70% — sensibly refusing to believe two attempts overturn a career's worth of context.
MAP blends prior knowledge with data, so two shots don't fool it; maximum likelihood, using data alone, is easily fooled.
MAP reports the most probable single value, not the average and not the uncertainty. The peak of a skewed distribution can sit far from where most of the probability is, so a MAP estimate alone can quietly mislead. If you need to know how sure you are, keep the whole posterior.