13.07.2015 Views

FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP

FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP

FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP

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.

109Appendix BAppendix to Section 4B.1 Pro<strong>of</strong>sPro<strong>of</strong> <strong>of</strong> Lemma 6: We first note thatP(α 0 ) . = e −nD(α 0/ρ||p) = e −n ρ D(α 0||ρp)(B.1)where as previously stated D(α 0 ||p) = ∑ f α 0,f log α 0,fp fis the information divergence (alsocalled the Kullback-Leibler distance) between α 0 and p. We use = . to denote exponential equality,i.e., we write f(n) = . e −nd logf(n)to denote lim = d and . .≤, ≥ are similarly defined. Inn→∞ nestablishing P(α 1 |α 0 ), we focus on a specific category f, and look to calculate(P |S ∩C f | = n ρ α 1,f | |C f | = n ρ α 0,f), (B.2)i.e., to calculate the probability that pruning introduces n ρ α 1,f elements, from C f to S, given thatthere are n ρ α 0,f elements <strong>of</strong> C f . Towards this we note that there is a total <strong>of</strong>|C f | = n ρ α 0,f (B.3)possible elements in C f which may be categorized, each with probability ɛ f , to belong to C 1 bythe categorization algorithm. The fraction <strong>of</strong> such elements that are asked to be categorized tobelong to C 1 , is defined by α to bex f := |S ∩C f||C f |=nρ α 1,f|C f |an event which happens with probability(P(x f ) = P |S ∩C f | = n ρ α 1,f | |C f | = n )ρ α 0,f= α 1,fα 0,f,(B.4).= e −|C f|I f (x f ) , (B.5)where in the above,I f (x f ) = x f log( x fɛ f)+(1−x f )log( 1−x f1−ɛ f) is the rate function <strong>of</strong> the binomialdistribution with parameter ɛ f (cf. [CT06]). Now given thatρ∏(P(α 1 |α 0 ) = P |S ∩C f | = n ρ α 1,f | |C f | = n )ρ α 0,f(B.6)f=1

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

Saved successfully!

Ooh no, something went wrong!