v2007.09.13 - Convex Optimization

5.4. EDM DEFINITION 3035.4.2.2.1 Example. List member constraints via Gram matrix.Capitalizing on identity (727) relating Gram and EDM D matrices, aconstraint set such astr ( − 1V DV e )2 ie T i = ‖xi ‖ 2⎫⎪tr ( ⎬− 1V DV (e 2 ie T j + e j e T i ) 2) 1 = xTi x jtr ( (733)− 1V DV e ) ⎪2 je T j = ‖xj ‖ 2 ⎭relates list member x i to x j to within an isometry through inner-productidentity [285,1-7]cos ψ ij =xT i x j‖x i ‖ ‖x j ‖For M list members, there are a total of M(M+1)/2 such constraints.(734)Consider the academic problem of finding a Gram matrix subject toconstraints on each and every entry of the corresponding EDM:findD∈S N h−V DV 1 2 ∈ SNsubject to 〈 D , (e i e T j + e j e T i ) 1 2〉= ď ij , i,j=1... N , i < j−V DV ≽ 0(735)where the ďij are given nonnegative constants. EDM D can, of course,be replaced with the equivalent Gram-form (717). Requiring only theself-adjointness property (1219) of the main-diagonal linear operator δ weget, for A∈ S N〈D , A〉 = 〈 δ(G)1 T + 1δ(G) T − 2G , A 〉 = 〈G , δ(A1) − A〉 2 (736)Then the problem equivalent to (735) becomes, for G∈ S N c ⇔ G1=0findG∈S N csubject toG ∈ S N〈G , δ ( (e i e T j + e j e T i )1 ) 〉− (e i e T j + e j e T i ) = ďij , i,j=1... N , i < jG ≽ 0 (737)