Active Learning for Parameter Estimation in Bayesian Networks

Proof Outline

n

Fairly complex. First prove lemma:

n

Let G(a) be the gamma function, Y(a) be the digamma function

G’(a)/G(a) and H be the entropy function and a_*=åa_j

n

Define:

d(a₁, … , a_r)

n

Then:

n

Now use function properties of Y to derive the final

result