The maximum-likelihood estimate, pulled by a prior. Give the unknown mean a belief before the data — a Gaussian N(μ₀, τ²) — and the point estimate becomes a precision-weighted blend of what you believed and what you saw. The same move on regression weights is ridge. Down the center, data flows in, the estimator computes, the estimate comes out. The blue team builds and defends it; the red team tries to break it.
source C. M. Bishop, Pattern Recognition and Machine Learning (Springer, 2006), §1.2.5 & §3.1.4 — no single stable primary link, marked AMBER. Rendered, not quoted.
Data x₁…xₙ ~ N(μ, σ²), σ² known, μ unknown. Prior μ ~ N(μ₀, τ²). The posterior is again Gaussian; its mode (the MAP) is closed-form:
Read it as precisions: data precision n/σ², prior precision 1/τ². The estimate is their weighted average of x̄ and μ₀ — so it always lands between the two.
Live weights for the current settings:
| source | precision | weight |
|---|
MAP = MLE + a prior. Take the log-likelihood of the-maximum-likelihood and add log-prior; the mode of that sum is the MAP. Flatten the prior (τ²→∞) and the term vanishes — MAP collapses back to MLE.
Regularization IS a prior. A Gaussian prior on weights makes the MAP objective equal the ridge objective; weight decay and the-reward-model's KL penalty are the same trick. Each is MAP wearing a different coat.
The blue team's live check: re-run every identity — flat-prior = MLE, the closed form, ridge equivalence, and prior wash-out — and confirm each holds. If red tampers, this badge is where it shows.
Two things go in. The evidence: a sample mean x̄ from n observations with noise σ² — this alone gives the MLE, x̄. The belief: a prior mean μ₀ and how sure you are of it, τ² (small τ² = stubborn prior; large τ² = open mind).
The estimator does not choose between them — it weighs them by precision. That is the whole game, and it is what you feed the panel below.
Every number is computed from the closed form on the spot — never looked up. Ridge readout: penalty λ = σ²/τ² = .
What the machine produces, proven: with a flat prior the MAP equals the MLE exactly; with the Gaussian prior it is the precision blend and always lies between μ₀ and x̄; the regression version is ridge exactly (λ = σ²/τ²); and as n→∞ the prior washes out — MAP → MLE. The current estimate is above; the identities are the output.
The blue team's witness (left) confirms these live; the red team (right) flips the sign of the prior to break the blend.
Full Bayes integrates over μ; MAP just finds the peak and reports its location. When the posterior is fat, sharp, or two-humped, that peak can sit where almost no mass is.
"MAP is the most probable value of the parameter." Cut. For continuous μ it maximizes a density; every single point has probability zero. It is the mode, not a probability.
"MAP and posterior-mean are the same answer." Cut. Equal only when the posterior is symmetric (as here, Gaussian). Skew them apart and mode ≠ mean.
"A prior is subjective fudge; regularization is principled." Kept, corrected. They are the same object: ridge = Gaussian prior, L1/lasso = Laplace prior, weight decay = MAP. The penalty is a −log-prior.
The red team's move: subtract the log-prior instead of adding it — flip its sign. Now the "prior" pushes the estimate away from μ₀, and the MAP no longer lands between μ₀ and x̄. The blue team's witness (window 7) is watching.
Flip the sign and the blend breaks — the estimate leaves the interval [μ₀, x̄], the witness recomputes, disagrees with the identities, and turns red. Nothing is faked; the attack is real and it is caught.