v2009.01.01 - Convex Optimization
v2009.01.01 - Convex Optimization v2009.01.01 - Convex Optimization
458 CHAPTER 6. CONE OF DISTANCE MATRICES The problem dual to maximum variance unfolding problem (1092) (less the Gram rank constraint) has been called the fastest mixing Markov process. Explored in [295], that dual has simple interpretations in graph and circuit theory and in mechanical and thermal systems. Optimal Gram rank turns out to be tightly bounded above by minimum multiplicity of the second smallest eigenvalue of a dual optimal variable. 6.5 EDM definition in 11 T Any EDM D corresponding to affine dimension r has representation D(V X ) ∆ = δ(V X V T X )1 T + 1δ(V X V T X ) T − 2V X V T X ∈ EDM N (1093) where R(V X ∈ R N×r )⊆ N(1 T ) = 1 ⊥ , VX T V X = δ 2 (VX T V X ) and V X is full-rank with orthogonal columns. (1094) Equation (1093) is simply the standard EDM definition (798) with a centered list X as in (880); Gram matrix X T X has been replaced with the subcompact singular value decomposition (A.6.2) 6.4 V X V T X ≡ V T X T XV ∈ S N c ∩ S N + (1095) This means: inner product VX TV X is an r×r diagonal matrix Σ of nonzero singular values. Vector δ(V X VX T ) may me decomposed into complementary parts by projecting it on orthogonal subspaces 1 ⊥ and R(1) : namely, P 1 ⊥( δ(VX V T X ) ) = V δ(V X V T X ) (1096) P 1 ( δ(VX V T X ) ) = 1 N 11T δ(V X V T X ) (1097) Of course δ(V X V T X ) = V δ(V X V T X ) + 1 N 11T δ(V X V T X ) (1098) ∆ 6.4 Subcompact SVD: V X VX T = Q √ Σ √ ΣQ T ≡ V T X T XV . So VX T is not necessarily XV (5.5.1.0.1), although affine dimension r = rank(VX T ) = rank(XV ). (937)
6.5. EDM DEFINITION IN 11 T 459 by (821). Substituting this into EDM definition (1093), we get the Hayden, Wells, Liu, & Tarazaga EDM formula [159,2] where D(V X , y) ∆ = y1 T + 1y T + λ N 11T − 2V X V T X ∈ EDM N (1099) λ ∆ = 2‖V X ‖ 2 F = 1 T δ(V X V T X )2 and y ∆ = δ(V X V T X ) − λ 2N 1 = V δ(V XV T X ) (1100) and y=0 if and only if 1 is an eigenvector of EDM D . Scalar λ becomes an eigenvalue when corresponding eigenvector 1 exists. 6.5 Then the particular dyad sum from (1099) y1 T + 1y T + λ N 11T ∈ S N⊥ c (1101) must belong to the orthogonal complement of the geometric center subspace (p.671), whereas V X V T X ∈ SN c ∩ S N + (1095) belongs to the positive semidefinite cone in the geometric center subspace. Proof. We validate eigenvector 1 and eigenvalue λ . (⇒) Suppose 1 is an eigenvector of EDM D . Then because it follows V T X 1 = 0 (1102) D1 = δ(V X V T X )1T 1 + 1δ(V X V T X )T 1 = N δ(V X V T X ) + ‖V X ‖ 2 F 1 ⇒ δ(V X V T X ) ∝ 1 (1103) For some κ∈ R + δ(V X V T X ) T 1 = N κ = tr(V T X V X ) = ‖V X ‖ 2 F ⇒ δ(V X V T X ) = 1 N ‖V X ‖ 2 F1 (1104) so y=0. (⇐) Now suppose δ(V X VX T)= λ 1 ; id est, y=0. Then 2N D = λ N 11T − 2V X V T X ∈ EDM N (1105) 1 is an eigenvector with corresponding eigenvalue λ . 6.5 e.g., when X = I in EDM definition (798).
- Page 407 and 408: 5.9. BRIDGE: CONVEX POLYHEDRA TO ED
- Page 409 and 410: 5.9. BRIDGE: CONVEX POLYHEDRA TO ED
- Page 411 and 412: 5.9. BRIDGE: CONVEX POLYHEDRA TO ED
- Page 413 and 414: 5.10. EDM-ENTRY COMPOSITION 413 of
- Page 415 and 416: 5.10. EDM-ENTRY COMPOSITION 415 The
- Page 417 and 418: 5.11. EDM INDEFINITENESS 417 5.11.1
- Page 419 and 420: 5.11. EDM INDEFINITENESS 419 we hav
- Page 421 and 422: 5.11. EDM INDEFINITENESS 421 So bec
- Page 423 and 424: 5.11. EDM INDEFINITENESS 423 where
- Page 425 and 426: 5.12. LIST RECONSTRUCTION 425 where
- Page 427 and 428: 5.12. LIST RECONSTRUCTION 427 (a) (
- Page 429 and 430: 5.13. RECONSTRUCTION EXAMPLES 429 D
- Page 431 and 432: 5.13. RECONSTRUCTION EXAMPLES 431 T
- Page 433 and 434: 5.13. RECONSTRUCTION EXAMPLES 433 w
- Page 435 and 436: 5.14. FIFTH PROPERTY OF EUCLIDEAN M
- Page 437 and 438: 5.14. FIFTH PROPERTY OF EUCLIDEAN M
- Page 439 and 440: 5.14. FIFTH PROPERTY OF EUCLIDEAN M
- Page 441 and 442: 5.14. FIFTH PROPERTY OF EUCLIDEAN M
- Page 443 and 444: 5.14. FIFTH PROPERTY OF EUCLIDEAN M
- Page 445 and 446: Chapter 6 Cone of distance matrices
- Page 447 and 448: 6.1. DEFINING EDM CONE 447 6.1 Defi
- Page 449 and 450: 6.2. POLYHEDRAL BOUNDS 449 This con
- Page 451 and 452: 6.3. √ EDM CONE IS NOT CONVEX 451
- Page 453 and 454: 6.4. A GEOMETRY OF COMPLETION 453 (
- Page 455 and 456: 6.4. A GEOMETRY OF COMPLETION 455 (
- Page 457: 6.4. A GEOMETRY OF COMPLETION 457 F
- Page 461 and 462: 6.5. EDM DEFINITION IN 11 T 461 6.5
- Page 463 and 464: 6.5. EDM DEFINITION IN 11 T 463 1 0
- Page 465 and 466: 6.5. EDM DEFINITION IN 11 T 465 6.5
- Page 467 and 468: 6.6. CORRESPONDENCE TO PSD CONE S N
- Page 469 and 470: 6.6. CORRESPONDENCE TO PSD CONE S N
- Page 471 and 472: 6.6. CORRESPONDENCE TO PSD CONE S N
- Page 473 and 474: 6.7. VECTORIZATION & PROJECTION INT
- Page 475 and 476: 6.7. VECTORIZATION & PROJECTION INT
- Page 477 and 478: 6.8. DUAL EDM CONE 477 When the Fin
- Page 479 and 480: 6.8. DUAL EDM CONE 479 Proof. First
- Page 481 and 482: 6.8. DUAL EDM CONE 481 EDM 2 = S 2
- Page 483 and 484: 6.8. DUAL EDM CONE 483 whose veraci
- Page 485 and 486: 6.8. DUAL EDM CONE 485 6.8.1.3.1 Ex
- Page 487 and 488: 6.8. DUAL EDM CONE 487 has dual aff
- Page 489 and 490: 6.8. DUAL EDM CONE 489 6.8.1.7 Scho
- Page 491 and 492: 6.9. THEOREM OF THE ALTERNATIVE 491
- Page 493 and 494: 6.10. POSTSCRIPT 493 When D is an E
- Page 495 and 496: Chapter 7 Proximity problems In sum
- Page 497 and 498: 497 project on the subspace, then p
- Page 499 and 500: 499 H S N h 0 EDM N K = S N h ∩ R
- Page 501 and 502: 501 7.0.3 Problem approach Problems
- Page 503 and 504: 7.1. FIRST PREVALENT PROBLEM: 503 f
- Page 505 and 506: 7.1. FIRST PREVALENT PROBLEM: 505 7
- Page 507 and 508: 7.1. FIRST PREVALENT PROBLEM: 507 d
6.5. EDM DEFINITION IN 11 T 459<br />
by (821). Substituting this into EDM definition (1093), we get the<br />
Hayden, Wells, Liu, & Tarazaga EDM formula [159,2]<br />
where<br />
D(V X , y) ∆ = y1 T + 1y T + λ N 11T − 2V X V T X ∈ EDM N (1099)<br />
λ ∆ = 2‖V X ‖ 2 F = 1 T δ(V X V T X )2 and y ∆ = δ(V X V T X ) − λ<br />
2N 1 = V δ(V XV T X )<br />
(1100)<br />
and y=0 if and only if 1 is an eigenvector of EDM D . Scalar λ becomes<br />
an eigenvalue when corresponding eigenvector 1 exists. 6.5<br />
Then the particular dyad sum from (1099)<br />
y1 T + 1y T + λ N 11T ∈ S N⊥<br />
c (1101)<br />
must belong to the orthogonal complement of the geometric center subspace<br />
(p.671), whereas V X V T X ∈ SN c ∩ S N + (1095) belongs to the positive semidefinite<br />
cone in the geometric center subspace.<br />
Proof. We validate eigenvector 1 and eigenvalue λ .<br />
(⇒) Suppose 1 is an eigenvector of EDM D . Then because<br />
it follows<br />
V T X 1 = 0 (1102)<br />
D1 = δ(V X V T X )1T 1 + 1δ(V X V T X )T 1 = N δ(V X V T X ) + ‖V X ‖ 2 F 1<br />
⇒ δ(V X V T X ) ∝ 1 (1103)<br />
For some κ∈ R +<br />
δ(V X V T X ) T 1 = N κ = tr(V T X V X ) = ‖V X ‖ 2 F ⇒ δ(V X V T X ) = 1 N ‖V X ‖ 2 F1 (1104)<br />
so y=0.<br />
(⇐) Now suppose δ(V X VX T)= λ 1 ; id est, y=0. Then<br />
2N<br />
D = λ N 11T − 2V X V T X ∈ EDM N (1105)<br />
1 is an eigenvector with corresponding eigenvalue λ . <br />
6.5 e.g., when X = I in EDM definition (798).