v2009.01.01 - Convex Optimization
v2009.01.01 - Convex Optimization v2009.01.01 - Convex Optimization
264 CHAPTER 4. SEMIDEFINITE PROGRAMMING problem (623) becomes equivalent to: (Theorem A.3.1.0.7) minimize 1 Tˆx X∈ S n , ˆx∈R n subject to A(ˆx + 1) 1 [ = b 2 ] X ˆx G = ˆx T 1 δ(X) = 1 (G ≽ 0) rankG = 1 (625) where solution is confined to rank-1 vertices of the elliptope in S n+1 (5.9.1.0.1) by the rank constraint, the positive semidefiniteness, and the equality constraints δ(X)=1. The rank constraint makes this problem nonconvex; by removing it 4.13 we get the semidefinite program minimize 1 Tˆx X∈ S n , ˆx∈R n subject to A(ˆx + 1) 1 [ = b 2 ] X ˆx G = ˆx T 1 δ(X) = 1 ≽ 0 (626) whose optimal solution x ⋆ (622) is identical to that of minimum cardinality Boolean problem (614) if and only if rankG ⋆ =1. Hope 4.14 of acquiring a rank-1 solution is not ill-founded because 2 n elliptope vertices have rank 1, and we are minimizing an affine function on a subset of the elliptope (Figure 106) containing rank-1 vertices; id est, by assumption that the feasible set of minimum cardinality Boolean problem (614) is nonempty, a desired solution resides on the elliptope relative boundary at a rank-1 vertex. 4.15 4.13 Relaxed problem (626) can also be derived via Lagrange duality; it is a dual of a dual program [sic] to (625). [263] [53,5, exer.5.39] [329,IV] [125,11.3.4] The relaxed problem must therefore be convex having a larger feasible set; its optimal objective value represents a generally loose lower bound (1563) on the optimal objective of problem (625). 4.14 A more deterministic approach to constraining rank and cardinality is developed in 4.6.0.0.9. 4.15 Confinement to the elliptope can be regarded as a kind of normalization akin to matrix A column normalization suggested in [101] and explored in Example 4.2.3.1.2.
4.2. FRAMEWORK 265 For the data given in (616), our semidefinite program solver (accurate in solution to approximately 1E-8) 4.16 finds optimal solution to (626) ⎡ round(G ⋆ ) = ⎢ ⎣ 1 1 1 −1 1 1 −1 1 1 1 −1 1 1 −1 1 1 1 −1 1 1 −1 −1 −1 −1 1 −1 −1 1 1 1 1 −1 1 1 −1 1 1 1 −1 1 1 −1 −1 −1 −1 1 −1 −1 1 ⎤ ⎥ ⎦ (627) near a rank-1 vertex of the elliptope in S n+1 ; its sorted eigenvalues, ⎡ λ(G ⋆ ) = ⎢ ⎣ 6.99999977799099 0.00000022687241 0.00000002250296 0.00000000262974 −0.00000000999738 −0.00000000999875 −0.00000001000000 ⎤ ⎥ ⎦ (628) The negative eigenvalues are undoubtedly finite-precision effects. Because the largest eigenvalue predominates by many orders of magnitude, we can expect to find a good approximation to a minimum cardinality Boolean solution by truncating all smaller eigenvalues. By so doing we find, indeed, ⎛⎡ x ⋆ = round ⎜⎢ ⎝⎣ 0.00000000127947 0.00000000527369 0.00000000181001 0.99999997469044 0.00000001408950 0.00000000482903 ⎤⎞ = e 4 (629) ⎥⎟ ⎦⎠ the desired result (617). 4.16 A typically ignored limitation of interior-point methods of solution is their relative accuracy of only about 1E-8 on a machine using 64-bit (double precision) floating-point arithmetic; id est, optimal solution x ⋆ cannot be more accurate than square root of machine epsilon (ǫ=2.2204E-16). Duality gap is not a good measure of solution accuracy.
- Page 213 and 214: 3.1. CONVEX FUNCTION 213 q(x) f(x)
- Page 215 and 216: 3.1. CONVEX FUNCTION 215 3.1.7.0.2
- Page 217 and 218: 3.1. CONVEX FUNCTION 217 We learned
- Page 219 and 220: 3.1. CONVEX FUNCTION 219 Since opti
- Page 221 and 222: 3.1. CONVEX FUNCTION 221 2 1.5 1 0.
- Page 223 and 224: 3.1. CONVEX FUNCTION 223 Setting th
- Page 225 and 226: 3.1. CONVEX FUNCTION 225 Similarly,
- Page 227 and 228: 3.1. CONVEX FUNCTION 227 For vector
- Page 229 and 230: 3.1. CONVEX FUNCTION 229 This means
- Page 231 and 232: 3.1. CONVEX FUNCTION 231 f(Y ) −
- Page 233 and 234: 3.2. MATRIX-VALUED CONVEX FUNCTION
- Page 235 and 236: 3.2. MATRIX-VALUED CONVEX FUNCTION
- Page 237 and 238: 3.2. MATRIX-VALUED CONVEX FUNCTION
- Page 239 and 240: 3.3. QUASICONVEX 239 exponential al
- Page 241 and 242: 3.3. QUASICONVEX 241 Unlike convex
- Page 243 and 244: 3.4. SALIENT PROPERTIES 243 6. (af
- Page 245 and 246: Chapter 4 Semidefinite programming
- Page 247 and 248: 4.1. CONIC PROBLEM 247 where K is a
- Page 249 and 250: 4.1. CONIC PROBLEM 249 4.1.1.2 Redu
- Page 251 and 252: 4.1. CONIC PROBLEM 251 In any SDP f
- Page 253 and 254: 4.1. CONIC PROBLEM 253 Proposition
- Page 255 and 256: 4.2. FRAMEWORK 255 sets are closed
- Page 257 and 258: 4.2. FRAMEWORK 257 4.2.1.1.3 Exampl
- Page 259 and 260: 4.2. FRAMEWORK 259 4.2.2 Duals The
- Page 261 and 262: 4.2. FRAMEWORK 261 When equality is
- Page 263: 4.2. FRAMEWORK 263 The pseudoinvers
- Page 267 and 268: 4.2. FRAMEWORK 267 minimizes an aff
- Page 269 and 270: 4.3. RANK REDUCTION 269 whose rank
- Page 271 and 272: 4.3. RANK REDUCTION 271 and where m
- Page 273 and 274: 4.3. RANK REDUCTION 273 4.3.3 Optim
- Page 275 and 276: 4.3. RANK REDUCTION 275 Initialize:
- Page 277 and 278: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 279 and 280: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 281 and 282: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 283 and 284: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 285 and 286: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 287 and 288: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 289 and 290: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 291 and 292: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 293 and 294: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 295 and 296: 4.5. CONSTRAINING CARDINALITY 295 m
- Page 297 and 298: 4.5. CONSTRAINING CARDINALITY 297 m
- Page 299 and 300: 4.5. CONSTRAINING CARDINALITY 299 a
- Page 301 and 302: 4.5. CONSTRAINING CARDINALITY 301 f
- Page 303 and 304: 4.5. CONSTRAINING CARDINALITY 303 n
- Page 305 and 306: 4.5. CONSTRAINING CARDINALITY 305 W
- Page 307 and 308: 4.5. CONSTRAINING CARDINALITY 307 t
- Page 309 and 310: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 311 and 312: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 313 and 314: 4.6. CARDINALITY AND RANK CONSTRAIN
264 CHAPTER 4. SEMIDEFINITE PROGRAMMING<br />
problem (623) becomes equivalent to: (Theorem A.3.1.0.7)<br />
minimize 1 Tˆx<br />
X∈ S n , ˆx∈R n<br />
subject to A(ˆx + 1) 1<br />
[ = b 2<br />
] X ˆx<br />
G =<br />
ˆx T 1<br />
δ(X) = 1<br />
(G ≽ 0)<br />
rankG = 1<br />
(625)<br />
where solution is confined to rank-1 vertices of the elliptope in S n+1<br />
(5.9.1.0.1) by the rank constraint, the positive semidefiniteness, and the<br />
equality constraints δ(X)=1. The rank constraint makes this problem<br />
nonconvex; by removing it 4.13 we get the semidefinite program<br />
minimize 1 Tˆx<br />
X∈ S n , ˆx∈R n<br />
subject to A(ˆx + 1) 1<br />
[ = b 2<br />
] X ˆx<br />
G =<br />
ˆx T 1<br />
δ(X) = 1<br />
≽ 0<br />
(626)<br />
whose optimal solution x ⋆ (622) is identical to that of minimum cardinality<br />
Boolean problem (614) if and only if rankG ⋆ =1. Hope 4.14 of acquiring a<br />
rank-1 solution is not ill-founded because 2 n elliptope vertices have rank 1,<br />
and we are minimizing an affine function on a subset of the elliptope<br />
(Figure 106) containing rank-1 vertices; id est, by assumption that the<br />
feasible set of minimum cardinality Boolean problem (614) is nonempty,<br />
a desired solution resides on the elliptope relative boundary at a rank-1<br />
vertex. 4.15<br />
4.13 Relaxed problem (626) can also be derived via Lagrange duality; it is a dual of a<br />
dual program [sic] to (625). [263] [53,5, exer.5.39] [329,IV] [125,11.3.4] The relaxed<br />
problem must therefore be convex having a larger feasible set; its optimal objective value<br />
represents a generally loose lower bound (1563) on the optimal objective of problem (625).<br />
4.14 A more deterministic approach to constraining rank and cardinality is developed in<br />
4.6.0.0.9.<br />
4.15 Confinement to the elliptope can be regarded as a kind of normalization akin to<br />
matrix A column normalization suggested in [101] and explored in Example 4.2.3.1.2.