v2010.10.26 - Convex Optimization
v2010.10.26 - Convex Optimization v2010.10.26 - Convex Optimization
312 CHAPTER 4. SEMIDEFINITE PROGRAMMINGI0S 2 +(confer Figure 99)∂H = {G | 〈G, I〉 = κ}Figure 84: Trace heuristic can be interpreted as minimization of a hyperplane,with normal I , over positive semidefinite cone drawn here in isometricallyisomorphic R 3 . Polar of direction vector W = I points toward origin.
4.4. RANK-CONSTRAINED SEMIDEFINITE PROGRAM 3134.4.1.2.1 Proof. Suppose 〈G ⋆ , W 〉= τ is satisfied for some nonnegativeconstant τ after any particular iteration (736) (1700a) of the twominimization problems. Once a particular value of τ is achieved, it can neverbe exceeded by subsequent iterations because existence of feasible G and Whaving that vector inner-product τ has been established simultaneously ineach problem. Because the infimum of vector inner-product of two positivesemidefinite matrix variables is zero, the nonincreasing sequence of iterationsis thus bounded below hence convergent because any bounded monotonicsequence in R is convergent. [258,1.2] [42,1.1] Local convergence to somenonnegative objective value τ is thereby established.Local convergence, in this context, means convergence to a fixed point ofpossibly infeasible rank. Only local convergence can be established becauseobjective 〈G , W 〉 , when instead regarded simultaneously in two variables(G , W ) , is generally multimodal. (3.8.0.0.3)Local convergence, convergence to τ ≠ 0 and definition of a stall, neverimplies nonexistence of a rank-n feasible solution to (736). A nonexistentrank-n feasible solution would mean certain failure to converge globally bydefinition (737) (convergence to τ ≠ 0) but, as proved, convex iteration alwaysconverges locally if not globally.When a rank-n feasible solution to (736) exists, it remains an openproblem to state conditions under which 〈G ⋆ , W ⋆ 〉=τ =0 (737) is achievedby iterative solution of semidefinite programs (736) and (1700a). ThenrankG ⋆ ≤ n and pair (G ⋆ , W ⋆ ) becomes a globally optimal fixed point ofiteration. There can be no proof of global convergence because of theimplicit high-dimensional multimodal manifold in variables (G , W). Whenstall occurs, direction vector W can be manipulated to steer out; e.g.,reversal of search direction as in Example 4.6.0.0.1, or reinitialization to arandom rank-(N − n) matrix in the same positive semidefinite cone face(2.9.2.3) demanded by the current iterate: given ordered diagonalizationG ⋆ = QΛQ T ∈ S N , then W = U ⋆ ΦU ⋆T where U ⋆ = Q(:, n+1:N)∈ R N×N−nand where eigenvalue vector λ(W) 1:N−n = λ(Φ) has nonnegative uniformlydistributed random entries in (0, 1] by selection of Φ∈ S N−n+ whileλ(W) N−n+1:N = 0. When this direction works, rank with respect to iterationtends to be noisily monotonic. Zero eigenvalues act as memory whilerandomness largely reduces likelihood of stall.
- Page 261 and 262: 3.6. GRADIENT 2613.6.4 second-order
- Page 263 and 264: 3.7. CONVEX MATRIX-VALUED FUNCTION
- Page 265 and 266: 3.7. CONVEX MATRIX-VALUED FUNCTION
- Page 267 and 268: 3.7. CONVEX MATRIX-VALUED FUNCTION
- Page 269 and 270: 3.8. QUASICONVEX 269exponential alw
- Page 271 and 272: 3.9. SALIENT PROPERTIES 2713.8.0.0.
- Page 273 and 274: Chapter 4Semidefinite programmingPr
- Page 275 and 276: 4.1. CONIC PROBLEM 275(confer p.162
- Page 277 and 278: 4.1. CONIC PROBLEM 277PCsemidefinit
- Page 279 and 280: 4.1. CONIC PROBLEM 279is the affine
- Page 281 and 282: 4.1. CONIC PROBLEM 281faces of S 3
- Page 283 and 284: 4.1. CONIC PROBLEM 2834.1.2.3 Previ
- Page 285 and 286: 4.2. FRAMEWORK 285Semidefinite Fark
- Page 287 and 288: 4.2. FRAMEWORK 287On the other hand
- Page 289 and 290: 4.2. FRAMEWORK 2894.2.2.1 Dual prob
- Page 291 and 292: 4.2. FRAMEWORK 291For symmetric pos
- Page 293 and 294: 4.2. FRAMEWORK 293has norm ‖x ⋆
- Page 295 and 296: 4.2. FRAMEWORK 295minimize 1 TˆxX
- Page 297 and 298: 4.2. FRAMEWORK 297asminimize ‖ỹ
- Page 299 and 300: 4.3. RANK REDUCTION 2994.3 Rank red
- Page 301 and 302: 4.3. RANK REDUCTION 301A rank-reduc
- Page 303 and 304: 4.3. RANK REDUCTION 303(t ⋆ i)
- Page 305 and 306: 4.3. RANK REDUCTION 3054.3.3.0.1 Ex
- Page 307 and 308: 4.3. RANK REDUCTION 3074.3.3.0.2 Ex
- Page 309 and 310: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 311: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 315 and 316: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 317 and 318: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 319 and 320: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 321 and 322: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 323 and 324: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 325 and 326: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 327 and 328: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 329 and 330: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 331 and 332: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 333 and 334: 4.5. CONSTRAINING CARDINALITY 333mi
- Page 335 and 336: 4.5. CONSTRAINING CARDINALITY 3350R
- Page 337 and 338: 4.5. CONSTRAINING CARDINALITY 337it
- Page 339 and 340: 4.5. CONSTRAINING CARDINALITY 339m/
- Page 341 and 342: 4.5. CONSTRAINING CARDINALITY 341we
- Page 343 and 344: 4.5. CONSTRAINING CARDINALITY 343fl
- Page 345 and 346: 4.5. CONSTRAINING CARDINALITY 345We
- Page 347 and 348: 4.5. CONSTRAINING CARDINALITY 3474.
- Page 349 and 350: 4.5. CONSTRAINING CARDINALITY 349R
- Page 351 and 352: 4.5. CONSTRAINING CARDINALITY 351pe
- Page 353 and 354: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 355 and 356: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 357 and 358: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 359 and 360: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 361 and 362: 4.6. CARDINALITY AND RANK CONSTRAIN
4.4. RANK-CONSTRAINED SEMIDEFINITE PROGRAM 3134.4.1.2.1 Proof. Suppose 〈G ⋆ , W 〉= τ is satisfied for some nonnegativeconstant τ after any particular iteration (736) (1700a) of the twominimization problems. Once a particular value of τ is achieved, it can neverbe exceeded by subsequent iterations because existence of feasible G and Whaving that vector inner-product τ has been established simultaneously ineach problem. Because the infimum of vector inner-product of two positivesemidefinite matrix variables is zero, the nonincreasing sequence of iterationsis thus bounded below hence convergent because any bounded monotonicsequence in R is convergent. [258,1.2] [42,1.1] Local convergence to somenonnegative objective value τ is thereby established.Local convergence, in this context, means convergence to a fixed point ofpossibly infeasible rank. Only local convergence can be established becauseobjective 〈G , W 〉 , when instead regarded simultaneously in two variables(G , W ) , is generally multimodal. (3.8.0.0.3)Local convergence, convergence to τ ≠ 0 and definition of a stall, neverimplies nonexistence of a rank-n feasible solution to (736). A nonexistentrank-n feasible solution would mean certain failure to converge globally bydefinition (737) (convergence to τ ≠ 0) but, as proved, convex iteration alwaysconverges locally if not globally.When a rank-n feasible solution to (736) exists, it remains an openproblem to state conditions under which 〈G ⋆ , W ⋆ 〉=τ =0 (737) is achievedby iterative solution of semidefinite programs (736) and (1700a). ThenrankG ⋆ ≤ n and pair (G ⋆ , W ⋆ ) becomes a globally optimal fixed point ofiteration. There can be no proof of global convergence because of theimplicit high-dimensional multimodal manifold in variables (G , W). Whenstall occurs, direction vector W can be manipulated to steer out; e.g.,reversal of search direction as in Example 4.6.0.0.1, or reinitialization to arandom rank-(N − n) matrix in the same positive semidefinite cone face(2.9.2.3) demanded by the current iterate: given ordered diagonalizationG ⋆ = QΛQ T ∈ S N , then W = U ⋆ ΦU ⋆T where U ⋆ = Q(:, n+1:N)∈ R N×N−nand where eigenvalue vector λ(W) 1:N−n = λ(Φ) has nonnegative uniformlydistributed random entries in (0, 1] by selection of Φ∈ S N−n+ whileλ(W) N−n+1:N = 0. When this direction works, rank with respect to iterationtends to be noisily monotonic. Zero eigenvalues act as memory whilerandomness largely reduces likelihood of stall.