10.03.2015 Views

Chapter 3 Geometry of convex functions - Meboo Publishing ...

Chapter 3 Geometry of convex functions - Meboo Publishing ...

Chapter 3 Geometry of convex functions - Meboo Publishing ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

238 CHAPTER 3. GEOMETRY OF CONVEX FUNCTIONS<br />

partitioned feasible sets are not interdependent, and the fact that the original<br />

problem (though nonlinear) is <strong>convex</strong> simultaneously in both variables. 3.14<br />

But partitioning alone does not guarantee a projector. To make<br />

orthogonal projector W a certainty, we must invoke a known analytical<br />

optimal solution to problem (553): Diagonalize optimal solution from<br />

problem (552) x ⋆ x ⋆T QΛQ T (A.5.1) and set U ⋆ = Q(:, 1:k)∈ R n×k<br />

per (1663c);<br />

W = U ⋆ U ⋆T = x⋆ x ⋆T<br />

‖x ⋆ ‖ 2 + Q(:, 2:k)Q(:, 2:k)T (554)<br />

Then optimal solution (x ⋆ , U ⋆ ) to problem (551) is found, for small ǫ , by<br />

iterating solution to problem (552) with optimal (projector) solution (554)<br />

to <strong>convex</strong> problem (553).<br />

Pro<strong>of</strong>. Optimal vector x ⋆ is orthogonal to the last n −1 columns <strong>of</strong><br />

orthogonal matrix Q , so<br />

f ⋆ (552) = ‖x ⋆ ‖ 2 (1 − (1 + ǫ) −1 ) (555)<br />

after each iteration. Convergence <strong>of</strong> f(552) ⋆ is proven with the observation that<br />

iteration (552) (553a) is a nonincreasing sequence that is bounded below by 0.<br />

Any bounded monotonic sequence in R is convergent. [252,1.2] [41,1.1]<br />

Expression (554) for optimal projector W holds at each iteration, therefore<br />

‖x ⋆ ‖ 2 (1 − (1 + ǫ) −1 ) must also represent the optimal objective value f(552)<br />

⋆<br />

at convergence.<br />

Because the objective f (551) from problem (551) is also bounded below<br />

by 0 on the same domain, this convergent optimal objective value f(552) ⋆ (for<br />

positive ǫ arbitrarily close to 0) is necessarily optimal for (551); id est,<br />

by (1646), and<br />

f ⋆ (552) ≥ f ⋆ (551) ≥ 0 (556)<br />

lim<br />

ǫ→0 +f⋆ (552) = 0 (557)<br />

Since optimal (x ⋆ , U ⋆ ) from problem (552) is feasible to problem (551), and<br />

because their objectives are equivalent for projectors by (548), then converged<br />

(x ⋆ , U ⋆ ) must also be optimal to (551) in the limit. Because problem (551)<br />

is <strong>convex</strong>, this represents a globally optimal solution.<br />

<br />

3.14 A <strong>convex</strong> problem has <strong>convex</strong> feasible set, and the objective surface has one and only<br />

one global minimum. [296, p.123]

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!