Maximum A Posteriori Estimation
MAP estimation chooses the mode of the posterior distribution. By Bayes’ theorem,
Here is the observed dataset, is the unknown parameter, is the likelihood, and is the prior. The posterior is the updated distribution over parameter values after seeing the data.
so the denominator can be ignored for optimization:
The evidence term does not depend on , so it changes the scale of the posterior but not the parameter value that maximizes it.
It is maximum likelihood plus a prior. In many ML objectives, the log-prior acts like regularization: a Gaussian prior on weights yields an penalty.
Worked scenario
A coin is tossed 10 times and lands heads 7 times. With a flat prior, the posterior mode is the observed head rate, , because the prior contributes no pull toward either side. With a symmetric prior, the prior behaves like one prior head and one prior tail for the posterior mode, giving .
A much stronger symmetric prior, , behaves like nine prior heads and nine prior tails for the mode, giving . The likelihood still favors heads, but the prior encodes a strong expectation that the coin is near fair, so the MAP estimate moves toward rather than staying at the sample proportion.
Caveats
MAP hides posterior uncertainty and can be sensitive to parameterization: a mode can move under nonlinear transformations. Bayesian statistics often reports posterior means, medians, intervals, or decisions rather than only the posterior mode.
References
Nav