v2010.10.26 - Convex Optimization
v2010.10.26 - Convex Optimization v2010.10.26 - Convex Optimization
216 CHAPTER 2. CONVEX GEOMETRYa point x from its coordinates t ⋆ (x) via (489), but does not prescribe theindex set I . There are at least two computational methods for specifying{Γj(i) ∗ } : one is combinatorial but sure to succeed, the other is a geometricmethod that searches for a minimum of a nonconvex function. We describethe latter:Convex problem (P)(P)maximize tt∈Rsubject to x − tv ∈ Kminimize λ T xλ∈R nsubject to λ T v = 1λ ∈ K ∗ (D) (490)is equivalent to definition (486) whereas convex problem (D) is its dual; 2.88meaning, primal and dual optimal objectives are equal t ⋆ = λ ⋆T x assumingSlater’s condition (p.285) is satisfied. Under this assumption of strongduality, λ ⋆T (x − t ⋆ v)= t ⋆ (1 − λ ⋆T v)=0; which implies, the primal problemis equivalent tominimize λ ⋆T (x − tv)t∈R(P) (491)subject to x − tv ∈ Kwhile the dual problem is equivalent tominimize λ T (x − t ⋆ v)λ∈R nsubject to λ T v = 1 (D) (492)λ ∈ K ∗Instead given coordinates t ⋆ (x) and a description of cone K , we proposeinversion by alternating solution of primal and dual problemsN∑minimize Γ ∗Tx∈R n i (x − t ⋆ iΓ i )i=1(493)subject to x − t ⋆ iΓ i ∈ K , i=1... N2.88 Form a Lagrangian associated with primal problem (P):L(t, λ) = t + λ T (x − tv) = λ T x + t(1 − λ T v) ,λ ≽K ∗ 0suptL(t, λ) = λ T x , 1 − λ T v = 0Dual variable (Lagrange multiplier [250, p.216]) λ generally has a nonnegative sense forprimal maximization with any cone membership constraint, whereas λ would have anonpositive sense were the primal instead a minimization problem having a membershipconstraint.
2.13. DUAL CONE & GENERALIZED INEQUALITY 217N∑minimize Γ ∗TΓ ∗ i (x ⋆ − t ⋆ iΓ i )i ∈Rn i=1subject to Γ ∗Ti Γ i = 1 ,Γ ∗ i ∈ K ∗ ,i=1... Ni=1... N(494)where dual extreme directions Γ ∗ i are initialized arbitrarily and ultimatelyascertained by the alternation. Convex problems (493) and (494) areiterated until convergence which is guaranteed by virtue of a monotonicallynonincreasing real sequence of objective values. Convergence can be fast.The mapping t ⋆ (x) is uniquely inverted when the necessarily nonnegativeobjective vanishes; id est, when Γ ∗Ti (x ⋆ − t ⋆ iΓ i )=0 ∀i. Here, a zeroobjective can occur only at the true solution x . But this global optimalitycondition cannot be guaranteed by the alternation because the commonobjective function, when regarded in both primal x and dual Γ ∗ i variablessimultaneously, is generally neither quasiconvex or monotonic. (3.8.0.0.3)Conversely, a nonzero objective at convergence is a certificate thatinversion was not performed properly. A nonzero objective indicates thatthe global minimum of a multimodal objective function could not be foundby this alternation. That is a flaw in this particular iterative algorithm forinversion; not in theory. 2.89 A numerical remedy is to reinitialize the Γ ∗ i todifferent values.2.89 The Proof 2.13.12.0.2, that suitable dual extreme directions {Γj ∗ } always exist, meansthat a global optimization algorithm would always find the zero objective of alternation(493) (494); hence, the unique inversion x. But such an algorithm can be combinatorial.
- Page 165 and 166: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 167 and 168: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 169 and 170: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 171 and 172: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 173 and 174: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 175 and 176: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 177 and 178: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 179 and 180: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 181 and 182: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 183 and 184: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 185 and 186: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 187 and 188: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 189 and 190: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 191 and 192: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 193 and 194: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 195 and 196: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 197 and 198: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 199 and 200: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 201 and 202: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 203 and 204: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 205 and 206: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 207 and 208: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 209 and 210: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 211 and 212: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 213 and 214: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 215: 2.13. DUAL CONE & GENERALIZED INEQU
- Page 219 and 220: Chapter 3Geometry of convex functio
- Page 221 and 222: 3.1. CONVEX FUNCTION 221f 1 (x)f 2
- Page 223 and 224: 3.1. CONVEX FUNCTION 223Rf(b)f(X
- Page 225 and 226: 3.2. PRACTICAL NORM FUNCTIONS, ABSO
- Page 227 and 228: 3.2. PRACTICAL NORM FUNCTIONS, ABSO
- Page 229 and 230: 3.2. PRACTICAL NORM FUNCTIONS, ABSO
- Page 231 and 232: 3.2. PRACTICAL NORM FUNCTIONS, ABSO
- Page 233 and 234: 3.2. PRACTICAL NORM FUNCTIONS, ABSO
- Page 235 and 236: 3.3. INVERTED FUNCTIONS AND ROOTS 2
- Page 237 and 238: 3.4. AFFINE FUNCTION 237rather]x >
- Page 239 and 240: 3.4. AFFINE FUNCTION 239f(z)Az 2z 1
- Page 241 and 242: 3.5. EPIGRAPH, SUBLEVEL SET 241{a T
- Page 243 and 244: 3.5. EPIGRAPH, SUBLEVEL SET 243Subl
- Page 245 and 246: 3.5. EPIGRAPH, SUBLEVEL SET 245wher
- Page 247 and 248: 3.5. EPIGRAPH, SUBLEVEL SET 247part
- Page 249 and 250: 3.5. EPIGRAPH, SUBLEVEL SET 249that
- Page 251 and 252: 3.6. GRADIENT 251respect to its vec
- Page 253 and 254: 3.6. GRADIENT 253Invertibility is g
- Page 255 and 256: 3.6. GRADIENT 2553.6.1.0.2 Theorem.
- Page 257 and 258: 3.6. GRADIENT 257f(Y )[ ∇f(X)−1
- Page 259 and 260: 3.6. GRADIENT 259αβα ≥ β ≥
- 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
2.13. DUAL CONE & GENERALIZED INEQUALITY 217N∑minimize Γ ∗TΓ ∗ i (x ⋆ − t ⋆ iΓ i )i ∈Rn i=1subject to Γ ∗Ti Γ i = 1 ,Γ ∗ i ∈ K ∗ ,i=1... Ni=1... N(494)where dual extreme directions Γ ∗ i are initialized arbitrarily and ultimatelyascertained by the alternation. <strong>Convex</strong> problems (493) and (494) areiterated until convergence which is guaranteed by virtue of a monotonicallynonincreasing real sequence of objective values. Convergence can be fast.The mapping t ⋆ (x) is uniquely inverted when the necessarily nonnegativeobjective vanishes; id est, when Γ ∗Ti (x ⋆ − t ⋆ iΓ i )=0 ∀i. Here, a zeroobjective can occur only at the true solution x . But this global optimalitycondition cannot be guaranteed by the alternation because the commonobjective function, when regarded in both primal x and dual Γ ∗ i variablessimultaneously, is generally neither quasiconvex or monotonic. (3.8.0.0.3)Conversely, a nonzero objective at convergence is a certificate thatinversion was not performed properly. A nonzero objective indicates thatthe global minimum of a multimodal objective function could not be foundby this alternation. That is a flaw in this particular iterative algorithm forinversion; not in theory. 2.89 A numerical remedy is to reinitialize the Γ ∗ i todifferent values.2.89 The Proof 2.13.12.0.2, that suitable dual extreme directions {Γj ∗ } always exist, meansthat a global optimization algorithm would always find the zero objective of alternation(493) (494); hence, the unique inversion x. But such an algorithm can be combinatorial.