v2010.10.26 - Convex Optimization
v2010.10.26 - Convex Optimization v2010.10.26 - Convex Optimization
484 CHAPTER 5. EUCLIDEAN DISTANCE MATRIX5.13.2.2 Isotonic solution with sort constraintBecause problems involving rank are generally difficult, we will partition(1147) into two problems we know how to solve and then alternate theirsolution until convergence:minimize ‖−VN T(D − O)V N ‖ FDsubject to rankVN TDV N ≤ 3D ∈ EDM Nminimize ‖σ − Πd‖σsubject to σ ∈ K M+(a)(b)(1149)where sort-index matrix O (a given constant in (a)) becomes an implicitvector variable o i solving the i th instance of (1149b)1√2dvec O i = o i Π T σ ⋆ ∈ R N(N−1)/2 , i∈{1, 2, 3...} (1150)As mentioned in discussion of relaxed problem (1148), a closed-formsolution to problem (1149a) exists. Only the first iteration of (1149a)sees the original sort-index matrix O whose entries are nonnegative wholenumbers; id est, O 0 =O∈ S N h ∩ R N×N+ (1144). Subsequent iterations i takethe previous solution of (1149b) as inputO i = dvec −1 ( √ 2o i ) ∈ S N (1151)real successors, estimating distance-square not order, to the sort-indexmatrix O .New convex problem (1149b) finds the unique minimum-distanceprojection of Πd on the monotone nonnegative cone K M+ . By definingY †T = [e 1 − e 2 e 2 −e 3 e 3 −e 4 · · · e m ] ∈ R m×m (430)where mN(N −1)/2, we may rewrite (1149b) as an equivalent quadraticprogram; a convex problem in terms of the halfspace-description of K M+ :minimize (σ − Πd) T (σ − Πd)σsubject to Y † σ ≽ 0(1152)
5.13. RECONSTRUCTION EXAMPLES 485This quadratic program can be converted to a semidefinite program viaSchur-form (3.5.2); we get the equivalent problemminimizet∈R , σsubject tot[tI σ − Πd(σ − Πd) T 1]≽ 0(1153)Y † σ ≽ 05.13.2.3 ConvergenceInE.10 we discuss convergence of alternating projection on intersectingconvex sets in a Euclidean vector space; convergence to a point in theirintersection. Here the situation is different for two reasons:Firstly, sets of positive semidefinite matrices having an upper bound onrank are generally not convex. Yet in7.1.4.0.1 we prove (1149a) is equivalentto a projection of nonincreasingly ordered eigenvalues on a subset of thenonnegative orthant:minimize ‖−VN T(D − O)V N ‖ FDsubject to rankVN TDV N ≤ 3D ∈ EDM N≡minimize ‖Υ − Λ‖ FΥ [ ]R3+subject to δ(Υ) ∈0(1154)where −VN TDV N UΥU T ∈ S N−1 and −VN TOV N QΛQ T ∈ S N−1 areordered diagonalizations (A.5). It so happens: optimal orthogonal U ⋆always equals Q given. Linear operator T(A) = U ⋆T AU ⋆ , acting on squarematrix A , is an isometry because Frobenius’ norm is orthogonally invariant(48). This isometric isomorphism T thus maps a nonconvex problem to aconvex one that preserves distance.Secondly, the second half (1149b) of the alternation takes place in adifferent vector space; S N h (versus S N−1 ). From5.6 we know these twovector spaces are related by an isomorphism, S N−1 =V N (S N h ) (1018), butnot by an isometry.We have, therefore, no guarantee from theory of alternating projectionthat the alternation (1149) converges to a point, in the set of allEDMs corresponding to affine dimension not in excess of 3, belonging todvec EDM N ∩ Π T K M+ .
- Page 433 and 434: 5.4. EDM DEFINITION 433because (A.3
- Page 435 and 436: 5.5. INVARIANCE 4355.5.1.0.1 Exampl
- Page 437 and 438: 5.5. INVARIANCE 437x 2 x 2x 3 x 1 x
- Page 439 and 440: 5.6. INJECTIVITY OF D & UNIQUE RECO
- Page 441 and 442: 5.6. INJECTIVITY OF D & UNIQUE RECO
- Page 443 and 444: 5.6. INJECTIVITY OF D & UNIQUE RECO
- Page 445 and 446: 5.7. EMBEDDING IN AFFINE HULL 4455.
- Page 447 and 448: 5.7. EMBEDDING IN AFFINE HULL 447Fo
- Page 449 and 450: 5.7. EMBEDDING IN AFFINE HULL 4495.
- Page 451 and 452: 5.8. EUCLIDEAN METRIC VERSUS MATRIX
- Page 453 and 454: 5.8. EUCLIDEAN METRIC VERSUS MATRIX
- Page 455 and 456: 5.8. EUCLIDEAN METRIC VERSUS MATRIX
- Page 457 and 458: 5.8. EUCLIDEAN METRIC VERSUS MATRIX
- Page 459 and 460: 5.9. BRIDGE: CONVEX POLYHEDRA TO ED
- Page 461 and 462: 5.9. BRIDGE: CONVEX POLYHEDRA TO ED
- Page 463 and 464: 5.9. BRIDGE: CONVEX POLYHEDRA TO ED
- Page 465 and 466: 5.10. EDM-ENTRY COMPOSITION 465of a
- Page 467 and 468: 5.10. EDM-ENTRY COMPOSITION 467Then
- Page 469 and 470: 5.11. EDM INDEFINITENESS 4695.11.1
- Page 471 and 472: 5.11. EDM INDEFINITENESS 471we have
- Page 473 and 474: 5.11. EDM INDEFINITENESS 473So beca
- Page 475 and 476: 5.11. EDM INDEFINITENESS 475holds o
- Page 477 and 478: 5.12. LIST RECONSTRUCTION 477where
- Page 479 and 480: 5.12. LIST RECONSTRUCTION 479(a)(c)
- Page 481 and 482: 5.13. RECONSTRUCTION EXAMPLES 481Wi
- Page 483: 5.13. RECONSTRUCTION EXAMPLES 483d
- Page 487 and 488: 5.14. FIFTH PROPERTY OF EUCLIDEAN M
- Page 489 and 490: 5.14. FIFTH PROPERTY OF EUCLIDEAN M
- Page 491 and 492: 5.14. FIFTH PROPERTY OF EUCLIDEAN M
- Page 493 and 494: 5.14. FIFTH PROPERTY OF EUCLIDEAN M
- Page 495 and 496: 5.14. FIFTH PROPERTY OF EUCLIDEAN M
- Page 497 and 498: Chapter 6Cone of distance matricesF
- Page 499 and 500: 6.1. DEFINING EDM CONE 4996.1 Defin
- Page 501 and 502: 6.2. POLYHEDRAL BOUNDS 501This cone
- Page 503 and 504: 6.4. EDM DEFINITION IN 11 T 503That
- Page 505 and 506: 6.4. EDM DEFINITION IN 11 T 505N(1
- Page 507 and 508: 6.4. EDM DEFINITION IN 11 T 507Then
- Page 509 and 510: 6.4. EDM DEFINITION IN 11 T 5096.4.
- Page 511 and 512: 6.5. CORRESPONDENCE TO PSD CONE S N
- Page 513 and 514: 6.5. CORRESPONDENCE TO PSD CONE S N
- Page 515 and 516: 6.5. CORRESPONDENCE TO PSD CONE S N
- Page 517 and 518: 6.6. VECTORIZATION & PROJECTION INT
- Page 519 and 520: 6.6. VECTORIZATION & PROJECTION INT
- Page 521 and 522: 6.6. VECTORIZATION & PROJECTION INT
- Page 523 and 524: 6.7. A GEOMETRY OF COMPLETION 523(b
- Page 525 and 526: 6.7. A GEOMETRY OF COMPLETION 525[3
- Page 527 and 528: 6.7. A GEOMETRY OF COMPLETION 527wh
- Page 529 and 530: 6.8. DUAL EDM CONE 529to the geomet
- Page 531 and 532: 6.8. DUAL EDM CONE 531Proof. First,
- Page 533 and 534: 6.8. DUAL EDM CONE 533EDM 2 = S 2 h
5.13. RECONSTRUCTION EXAMPLES 485This quadratic program can be converted to a semidefinite program viaSchur-form (3.5.2); we get the equivalent problemminimizet∈R , σsubject tot[tI σ − Πd(σ − Πd) T 1]≽ 0(1153)Y † σ ≽ 05.13.2.3 ConvergenceInE.10 we discuss convergence of alternating projection on intersectingconvex sets in a Euclidean vector space; convergence to a point in theirintersection. Here the situation is different for two reasons:Firstly, sets of positive semidefinite matrices having an upper bound onrank are generally not convex. Yet in7.1.4.0.1 we prove (1149a) is equivalentto a projection of nonincreasingly ordered eigenvalues on a subset of thenonnegative orthant:minimize ‖−VN T(D − O)V N ‖ FDsubject to rankVN TDV N ≤ 3D ∈ EDM N≡minimize ‖Υ − Λ‖ FΥ [ ]R3+subject to δ(Υ) ∈0(1154)where −VN TDV N UΥU T ∈ S N−1 and −VN TOV N QΛQ T ∈ S N−1 areordered diagonalizations (A.5). It so happens: optimal orthogonal U ⋆always equals Q given. Linear operator T(A) = U ⋆T AU ⋆ , acting on squarematrix A , is an isometry because Frobenius’ norm is orthogonally invariant(48). This isometric isomorphism T thus maps a nonconvex problem to aconvex one that preserves distance.Secondly, the second half (1149b) of the alternation takes place in adifferent vector space; S N h (versus S N−1 ). From5.6 we know these twovector spaces are related by an isomorphism, S N−1 =V N (S N h ) (1018), butnot by an isometry.We have, therefore, no guarantee from theory of alternating projectionthat the alternation (1149) converges to a point, in the set of allEDMs corresponding to affine dimension not in excess of 3, belonging todvec EDM N ∩ Π T K M+ .