v2010.10.26 - Convex Optimization
v2010.10.26 - Convex Optimization v2010.10.26 - Convex Optimization
576 CHAPTER 7. PROXIMITY PROBLEMSUnidimensional scaling, [105] a historically practical application ofmultidimensional scaling (5.12), entails solution of an optimization problemhaving local minima whose multiplicity varies as the factorial of point-listcardinality; geometrically, it means reconstructing a list constrained to lie inone affine dimension. In terms of point list, the nonconvex problem is: givennonnegative symmetric matrix H = [h ij ] ∈ S N ∩ R N×N+ (1357) whose entriesh ij are all known,minimize{x i ∈R}N∑(|x i − x j | − h ij ) 2 (1308)i , j=1called a raw stress problem [53, p.34] which has an implicit constraint ondimensional embedding of points {x i ∈ R , i=1... N}. This problem hasproven NP-hard; e.g., [73].As always, we first transform variables to distance-square D ∈ S N h ; sobegin with convex problem (1359) on page 569minimize − tr(V (D − 2Y )V )D , Y[ ]dij y ijsubject to≽ 0 ,y ijh 2 ijY ∈ S N hD ∈ EDM NrankV T N DV N = 1N ≥ j > i = 1... N −1(1378)that becomes equivalent to (1308) by making explicit the constraint on affinedimension via rank. The iteration is formed by moving the dimensionalconstraint to the objective:minimize −〈V (D − 2Y )V , I〉 − w〈VN TDV N , W 〉D , Y[ ]dij y ijsubject to≽ 0 , N ≥ j > i = 1... N −1y ij h 2 ij(1379)Y ∈ S N hD ∈ EDM Nwhere w (≈ 10) is a positive scalar just large enough to make 〈V T N DV N , W 〉vanish to within some numerical precision, and where direction matrix W is
7.2. SECOND PREVALENT PROBLEM: 577an optimal solution to semidefinite program (1700a)minimize −〈VN T WD⋆ V N , W 〉subject to 0 ≼ W ≼ ItrW = N − 1(1380)one of which is known in closed form. Semidefinite programs (1379) and(1380) are iterated until convergence in the sense defined on page 309. Thisiteration is not a projection method. (4.4.1.1) Convex problem (1379)is neither a relaxation of unidimensional scaling problem (1378); instead,problem (1379) is a convex equivalent to (1378) at convergence of theiteration.Jan de Leeuw provided us with some test data⎡H =⎢⎣0.000000 5.235301 5.499274 6.404294 6.486829 6.2632655.235301 0.000000 3.208028 5.840931 3.559010 5.3534895.499274 3.208028 0.000000 5.679550 4.020339 5.2398426.404294 5.840931 5.679550 0.000000 4.862884 4.5431206.486829 3.559010 4.020339 4.862884 0.000000 4.6187186.263265 5.353489 5.239842 4.543120 4.618718 0.000000and a globally optimal solution⎤⎥⎦(1381)X ⋆ = [ −4.981494 −2.121026 −1.038738 4.555130 0.764096 2.822032 ]= [ x ⋆ 1 x ⋆ 2 x ⋆ 3 x ⋆ 4 x ⋆ 5 x ⋆ 6 ](1382)found by searching 6! local minima of (1308) [105]. By iterating convexproblems (1379) and (1380) about twenty times (initial W = 0) we find theglobal infimum 98.12812 of stress problem (1308), and by (1132) we find acorresponding one-dimensional point list that is a rigid transformation in Rof X ⋆ .Here we found the infimum to accuracy of the given data, but that ceasesto hold as problem size increases. Because of machine numerical precisionand an interior-point method of solution, we speculate, accuracy degradesquickly as problem size increases beyond this.
- 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
- Page 535 and 536: 6.8. DUAL EDM CONE 535therefore the
- Page 537 and 538: 6.8. DUAL EDM CONE 537Elegance of t
- Page 539 and 540: 6.8. DUAL EDM CONE 5396.8.1.5 Affin
- Page 541 and 542: 6.8. DUAL EDM CONE 5416.8.1.7 Schoe
- Page 543 and 544: 6.8. DUAL EDM CONE 5430dvec rel ∂
- Page 545 and 546: 6.10. POSTSCRIPT 5456.10 Postscript
- Page 547 and 548: Chapter 7Proximity problemsIn the
- Page 549 and 550: 549project on the subspace, then pr
- Page 551 and 552: 551HS N h0EDM NK = S N h ∩ R N×N
- Page 553 and 554: 5537.0.3 Problem approachProblems t
- Page 555 and 556: 7.1. FIRST PREVALENT PROBLEM: 555fi
- Page 557 and 558: 7.1. FIRST PREVALENT PROBLEM: 5577.
- Page 559 and 560: 7.1. FIRST PREVALENT PROBLEM: 559di
- Page 561 and 562: 7.1. FIRST PREVALENT PROBLEM: 5617.
- Page 563 and 564: 7.1. FIRST PREVALENT PROBLEM: 563wh
- Page 565 and 566: 7.1. FIRST PREVALENT PROBLEM: 565Th
- Page 567 and 568: 7.2. SECOND PREVALENT PROBLEM: 567O
- Page 569 and 570: 7.2. SECOND PREVALENT PROBLEM: 569S
- Page 571 and 572: 7.2. SECOND PREVALENT PROBLEM: 571r
- Page 573 and 574: 7.2. SECOND PREVALENT PROBLEM: 573w
- Page 575: 7.2. SECOND PREVALENT PROBLEM: 5757
- Page 579 and 580: 7.3. THIRD PREVALENT PROBLEM: 579is
- Page 581 and 582: 7.3. THIRD PREVALENT PROBLEM: 581We
- Page 583 and 584: 7.3. THIRD PREVALENT PROBLEM: 583su
- Page 585 and 586: 7.3. THIRD PREVALENT PROBLEM: 585Gi
- Page 587 and 588: 7.3. THIRD PREVALENT PROBLEM: 587Op
- Page 589 and 590: 7.4. CONCLUSION 589filtering, multi
- Page 591 and 592: Appendix ALinear algebraA.1 Main-di
- Page 593 and 594: A.1. MAIN-DIAGONAL δ OPERATOR, λ
- Page 595 and 596: A.1. MAIN-DIAGONAL δ OPERATOR, λ
- Page 597 and 598: A.2. SEMIDEFINITENESS: DOMAIN OF TE
- Page 599 and 600: A.3. PROPER STATEMENTS 599(AB) T
- Page 601 and 602: A.3. PROPER STATEMENTS 601A.3.1Semi
- Page 603 and 604: A.3. PROPER STATEMENTS 603For A dia
- Page 605 and 606: A.3. PROPER STATEMENTS 605Diagonali
- Page 607 and 608: A.3. PROPER STATEMENTS 607For A,B
- Page 609 and 610: A.3. PROPER STATEMENTS 609When B is
- Page 611 and 612: A.4. SCHUR COMPLEMENT 611A.4 Schur
- Page 613 and 614: A.4. SCHUR COMPLEMENT 613A.4.0.0.3
- Page 615 and 616: A.4. SCHUR COMPLEMENT 615From Corol
- Page 617 and 618: A.5. EIGENVALUE DECOMPOSITION 617wh
- Page 619 and 620: A.5. EIGENVALUE DECOMPOSITION 619A.
- Page 621 and 622: A.6. SINGULAR VALUE DECOMPOSITION,
- Page 623 and 624: A.6. SINGULAR VALUE DECOMPOSITION,
- Page 625 and 626: A.7. ZEROS 625A.6.5SVD of symmetric
576 CHAPTER 7. PROXIMITY PROBLEMSUnidimensional scaling, [105] a historically practical application ofmultidimensional scaling (5.12), entails solution of an optimization problemhaving local minima whose multiplicity varies as the factorial of point-listcardinality; geometrically, it means reconstructing a list constrained to lie inone affine dimension. In terms of point list, the nonconvex problem is: givennonnegative symmetric matrix H = [h ij ] ∈ S N ∩ R N×N+ (1357) whose entriesh ij are all known,minimize{x i ∈R}N∑(|x i − x j | − h ij ) 2 (1308)i , j=1called a raw stress problem [53, p.34] which has an implicit constraint ondimensional embedding of points {x i ∈ R , i=1... N}. This problem hasproven NP-hard; e.g., [73].As always, we first transform variables to distance-square D ∈ S N h ; sobegin with convex problem (1359) on page 569minimize − tr(V (D − 2Y )V )D , Y[ ]dij y ijsubject to≽ 0 ,y ijh 2 ijY ∈ S N hD ∈ EDM NrankV T N DV N = 1N ≥ j > i = 1... N −1(1378)that becomes equivalent to (1308) by making explicit the constraint on affinedimension via rank. The iteration is formed by moving the dimensionalconstraint to the objective:minimize −〈V (D − 2Y )V , I〉 − w〈VN TDV N , W 〉D , Y[ ]dij y ijsubject to≽ 0 , N ≥ j > i = 1... N −1y ij h 2 ij(1379)Y ∈ S N hD ∈ EDM Nwhere w (≈ 10) is a positive scalar just large enough to make 〈V T N DV N , W 〉vanish to within some numerical precision, and where direction matrix W is