26.10.2013 Views

Nonparametric Bayesian Discrete Latent Variable Models for ...

Nonparametric Bayesian Discrete Latent Variable Models for ...

Nonparametric Bayesian Discrete Latent Variable Models for ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

increment process with the corresponding Lévy measure given by<br />

3.1 The Dirichlet Process<br />

dN(x) = x −1 e −x/β αdx, (3.10)<br />

where α and β are positive scalars, t ∈ [0, 1] and Q(t) is a distribution function on [0, 1].<br />

That is, defining the distribution of the random jump heights J1 ≥ J2, . . . to be<br />

P (J1 ≤ x1) = exp N(x1) <br />

P (Jj ≤ xj | Jj−1 = xj−1, . . . , J1 = x1) = exp (3.11)<br />

N(xj) − N(xj−1)<br />

and the random variables related to the jump times U1, U2, . . . to be i.i.d. Uni<strong>for</strong>m(0, 1),<br />

independent from J1, J2, . . . , the gamma process Zt is defined as<br />

Zt =<br />

∞<br />

JjI Uj ∈ 0, Q(t) . (3.12)<br />

j=1<br />

See Ferguson and Klass (1972) and Ferguson (1973) <strong>for</strong> details.<br />

Let Γ (1) ≥ Γ (2) ≥ . . . denote the jump heights of a gamma process with the scale<br />

parameter β = 1. And let θk ∼ G0 i.i.d., independent also of the Γ (k). The random<br />

measure defined as<br />

has a DP(α, G0) distribution where<br />

G =<br />

Pk =<br />

∞<br />

k=1<br />

Pkδθk (·) (3.13)<br />

Γ (k)<br />

∞ j=1 Γ . (3.14)<br />

(j)<br />

This definition explicitly shows the discreteness of the DP as it expresses G as an<br />

infinite sum of atomic measures. The practical limitation of this construction is that we<br />

need to know the value of the infinite sum to know the value of any of the weights. The<br />

next section summarizes another infinite sum representation of the DP which does not<br />

require evaluating the infinite sum, and allows approximating the DP by truncation.<br />

3.1.4 Stick Breaking Construction<br />

Sethuraman and Tiwari (1982) proposed a constructive definition of the DP based on a<br />

sequence of i.i.d. random variables. The proof is provided by Sethuraman (1994). Let<br />

vk be i.i.d. with a common distribution vk ∼ Beta(1, α). Define<br />

πk = vk<br />

j=1<br />

k−1 <br />

(1 − vj), (3.15)<br />

and let θk be independent of the vk and i.i.d. among themselves with common distribution<br />

G0. The random probability measure G that puts weights πk at the degenerate<br />

17

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!