24.02.2013 Views

Optimality

Optimality

Optimality

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.

286 V. J. Ribeiro, R. H. Riedi and R. G. Baraniuk<br />

we have<br />

(7.29)<br />

cov(Vγk, E)<br />

cov(Vγ, E) =<br />

⎛<br />

� �1/2 var(Vγk)<br />

var(Vγ)<br />

�<br />

var(Vγk)<br />

=: ξγ,k≥<br />

var(Vγ)<br />

⎝ ϱ2 γk<br />

� 1/2<br />

+ var(Wγk)<br />

var(Vγ)<br />

ϱ 2 γk<br />

From (7.30) we see that ξγ,k is not a function of E.<br />

Denote the covariance between Vγ and leaf node vector L = [ℓi]∈Λγk(n) as<br />

Θγ,L = [cov(Vγ, ℓi)] T . Then (7.30) gives<br />

(7.30) Θγk,L = ξγ,kΘγ,L.<br />

From (4.2) we have<br />

(7.31) E(Vγ|L) = var(Vγ)−ϕ(γ, L)<br />

where ϕ(γ, L) = ΘT γ,LQ−1 L Θγ,L. Note that ϕ(γ, L)≥0 since Q −1<br />

L is positive semidefinite.<br />

Using (7.30) we similarly get<br />

(7.32) E(Vγk|L) = var(Vγk)−<br />

.<br />

ϕ(γ, L)<br />

ξ2 .<br />

γ,k<br />

From (7.31) and (7.32) we see thatE(Vγ|L) andE(Vγk|L) are both minimized over<br />

L∈Λγk(n) by the same leaf vector that maximizes ϕ(γ, L). This proves Claim (2).<br />

Claim (3): µγ,γk(n) is a positive, non-decreasing, and discrete-concave function<br />

of n,∀k, γ.<br />

We start at a node γ at one scale from the bottom of the tree and then move up<br />

the tree.<br />

Initial Condition: Note that Vγk is a leaf node. From (2.1) and (??) we obtain<br />

(7.33) E(Vγ|Vγk) = var(Vγ)−<br />

(ϱγkvar(Vγ)) 2<br />

var(Vγk)<br />

⎞<br />

⎠<br />

≤ var(Vγ).<br />

For our choice of γ, µγ,γk(1) corresponds toE(Vγ|Vγk) −1 and µγ,γk(0) corresponds<br />

to 1/var(Vγ). Thus from (7.33), µγ,γk(n) is positive, non-decreasing, and discreteconcave<br />

(trivially since n takes only two values here).<br />

Induction Step: Given that µγ,γk(n) is a positive, non-decreasing, and discreteconcave<br />

function of n for k = 1, . . . , Pγ, we prove the same when γ is replaced by<br />

γ↑. Without loss of generality choose k such that (γ↑)k = γ. From (3.11), (3.13),<br />

(7.31), (7.32) and Claim (2), we have for L∈Lγ(n)<br />

(7.34)<br />

µγ(n) =<br />

µγ↑,k(n) =<br />

1<br />

var(Vγ) ·<br />

1<br />

var(Vγ↑) ·<br />

1<br />

1− ϕ(γ,L)<br />

1−<br />

var(Vγ)<br />

1<br />

ϕ(γ,L)<br />

, and<br />

ξ 2<br />

γ↑,k var(Vγ↑)<br />

From (7.26), the assumption that µγ,γk(n)∀k is a positive, non-decreasing, and<br />

discrete-concave function of n, and Lemma 3.1 we have that µγ(n) is a nondecreasing<br />

and discrete-concave function of n. Note that by definition (see (3.11))<br />

.<br />

1/2

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

Saved successfully!

Ooh no, something went wrong!