MCMC, Gibbs Örneklemleri

State \(x\):

\(x = (x_1, x_2, \ldots, x_i, \ldots, x_n)\)

State \(x'\):

\(x' = (x_1, x_2, \ldots, x'_i, \ldots, x_n)\)

Note These two states are identical except for coordinate \(i\):

Therefore: \(x_{-i} = x'_{-i}\) (they share all the same non-\(i\) coordinates)

Theorem

If \(\exists\) distribution \(\mu\) such that \(\mu(x) T(x \to x') = \mu(x') T(x' \to x)\), then \(\mu\) is stationary.

given \(T(x \to x') = p(x'_i \mid x_{-i})\)

LHS in English

\(p(x) \cdot p(x'_i \mid x_{-i})\)

\(p(x)\): “The probability of being in state \(x\)” (our candidate \(\mu(x)\) )

This is the full joint: \(\mu(x_1, x_2, \ldots, x_i, \ldots, x_n)\)

\(p(x'_i \mid x_{-i})\): “The probability of sampling value \(x'_i\) for variable \(i\), given all other variables are at \(x_{-i}\)

This is the Gibbs transition: we’re sampling a new value for coordinate \(i\)

Together: The probability we’re currently at state \(x\), multiplied by the probability the Gibbs sampler transitions us to state \(x'\) (by sampling \(x'_i\) for the \(i\)-th coordinate)

Yukarı