THE CONJUGATE PRIOR

A prior family closed under evidence. Put a Beta belief on a coin's bias, observe heads and tails, and the posterior is another Beta — belief update collapses to arithmetic: Beta(α+h, β+t). Rendered, not quoted.

SOURCE Raiffa & Schlaifer, Applied Statistical Decision Theory (1961), Harvard Business School — the founding text of conjugate analysis. archive.org/details/appliedstatistic00raif AMBER · book, no per-page anchor

Blue Team · builds & defends
3

The Model

Bias θ of a coin is unknown. Belief is a Beta(α,β) density on [0,1]. Data are h heads and t tails (n = h+t), each toss Bernoulli(θ).

prior  ∝ θα-1(1-θ)β-1
like   ∝ θh(1-θ)t

Multiply: exponents simply add. Beta × Bernoulli lands back in Beta. That is conjugacy — the family is closed under the likelihood.

5

The Lineage

the-bayes gave the rule posterior ∝ prior × likelihood — true always, tractable rarely. The conjugate prior makes it closed-form: no integral for the normaliser, update is (α,β) → (α+h, β+t).

This algebra is the engine beneath online Bayesian learning and Thompson sampling — a bandit that keeps one Beta per arm and updates it per pull.

7

The Witness

Live re-check: the panel's posterior must equal the closed form Beta(α+h,β+t), and its mean must lie between prior mean and sample mean. Runs on every change and after any tamper.

witness idle
The Machine
4

Data In IN ↓

2
2
7
3

prior mean = α/(α+β)  ·  sample mean = h/n

↓  MULTIPLY & NORMALISE  ↓
0

The Panel LIT

prior likelihood posterior
posterior = Beta(α+h, β+t)

mean is a convex blend of prior & data

↓  PROVEN RESULT  ↓
8

Data Out OUT ↓

booting…

Laplace check pending

Red Team · attacks & breaks
1

The Adversary

WALL Conjugacy is a convenience, not a law of nature. If the true prior is bimodal (a two-headed vs. fair coin), no single Beta represents it — the closed form buys tractability by constraining the belief shape. Model misspecification hides inside the algebra.

It also assumes exchangeable, i.i.d. tosses. Drift, streaks, or a coin that wears mid-experiment break the Bernoulli likelihood, and the tidy update quietly encodes a false model.

2

The Graveyard

  • "The posterior mean is just the sample mean h/n."
    No — it is a convex pull of the sample mean toward the prior mean; weight (α+β)/(α+β+n) is the prior's pseudo-count.
  • "More data can't hurt a Beta prior."
    A stubborn prior (huge α,β) needs large n to move; the same conjugacy that speeds updates also encodes stubbornness.
  • "Batch and one-at-a-time give different answers."
    They are identical — addition is associative. The panel proves it live.
6

The Tamper

Planted void: swap the counts — add heads to β and tails to α. The posterior slides the wrong way and its mean escapes the [prior, sample] interval. The Witness (7) catches it.