10.03.2015 Views

v2009.01.01 - Convex Optimization

v2009.01.01 - Convex Optimization

v2009.01.01 - Convex Optimization

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.

5.10. EDM-ENTRY COMPOSITION 413<br />

of any triangle. Because the smallest eigenvalue is 0, affine dimension r of<br />

any point list corresponding to D cannot exceed N −2. (5.7.1.1) <br />

5.10 EDM-entry composition<br />

Laurent [202,2.3] applies results from Schoenberg (1938) [271] to show<br />

certain nonlinear compositions of individual EDM entries yield EDMs;<br />

in particular,<br />

D ∈ EDM N ⇔ [1 − e −αd ij<br />

] ∈ EDM N ∀α > 0<br />

⇔ [e −αd ij<br />

] ∈ E N ∀α > 0<br />

(997)<br />

where D = [d ij ] and E N is the elliptope (981).<br />

5.10.0.0.1 Proof. (Laurent, 2003) [271] (confer [197])<br />

Lemma 2.1. from A Tour d’Horizon ...on Completion Problems. [202]<br />

The following assertions are equivalent: for D=[d ij , i,j=1... N]∈ S N h and<br />

E N the elliptope in S N (5.9.1.0.1),<br />

(i) D ∈ EDM N<br />

(ii) e −αD ∆ = [e −αd ij<br />

] ∈ E N for all α > 0<br />

(iii) 11 T − e −αD ∆ = [1 − e −αd ij<br />

] ∈ EDM N for all α > 0<br />

⋄<br />

1) Equivalence of Lemma 2.1 (i) (ii) is stated in Schoenberg’s Theorem 1<br />

[271, p.527].<br />

2) (ii) ⇒ (iii) can be seen from the statement in the beginning of section 3,<br />

saying that a distance space embeds in L 2 iff some associated matrix<br />

is PSD. We reformulate it:<br />

Let d =(d ij ) i,j=0,1...N be a distance space on N+1 points<br />

(i.e., symmetric hollow matrix of order N+1) and let<br />

p =(p ij ) i,j=1...N be the symmetric matrix of order N related by:<br />

(A) 2p ij = d 0i + d 0j − d ij for i,j = 1... N<br />

or equivalently<br />

(B) d 0i = p ii , d ij = p ii + p jj − 2p ij for i,j = 1... N

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

Saved successfully!

Ooh no, something went wrong!