08.10.2016 Views

Foundations of Data Science

2dLYwbK

2dLYwbK

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Probability q(p) <strong>of</strong><br />

the root being 1 1/2<br />

as a function <strong>of</strong> p<br />

1<br />

0<br />

high<br />

temperature<br />

low<br />

temperature<br />

0 1/2 1<br />

Probability p <strong>of</strong> a leaf being 1<br />

at phase transition<br />

slope <strong>of</strong> q(p) equals 1<br />

at p = 1/2<br />

Figure 9.10: Shape <strong>of</strong> q as a function <strong>of</strong> p for the height one tree and three values <strong>of</strong> β corresponding<br />

to low temperature, the phase transition temperature, and high temperature.<br />

.<br />

decreasing. From this it follows that q is monotonically increasing.<br />

In the iteration going from p to q, we do not get the true marginal probabilities at<br />

each level since we ignored the effect <strong>of</strong> the portion <strong>of</strong> the tree above. However, when we<br />

get to the root, we do get the true marginal for the root. To get the true marginal’s for<br />

the interior nodes we need to send messages down from the root.<br />

Note: The joint probability distribution for the tree is <strong>of</strong> the form e β ∑<br />

(ij)∈E)<br />

x i x j ∏ = e βx ix j<br />

.<br />

(i,j)∈E<br />

Suppose x 1 has value 1 with probability p. Then define a function ϕ, called evidence, such<br />

that<br />

{ p for x1 = 1<br />

ϕ (x 1 ) =<br />

1 − p for x 1 = −1<br />

= ( p − 1 2)<br />

x1 + 1 2<br />

and multiply the joint probability function by ϕ. Note, however, that the marginal probability<br />

<strong>of</strong> x 1 is not p. In fact, it may be further from p after multiplying the conditional<br />

probability function by the function ϕ.<br />

326

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

Saved successfully!

Ooh no, something went wrong!