10.03.2015 Views

v2009.01.01 - Convex Optimization

v2009.01.01 - Convex Optimization

v2009.01.01 - Convex Optimization

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

318 CHAPTER 4. SEMIDEFINITE PROGRAMMING<br />

The trace constraint on X normalizes vector x while the diagonal constraint<br />

on Z maintains sign between respective entries of x and y . Regularization<br />

term ‖X −Y ‖ F then makes x equal to y to within a real scalar; (C.2.0.0.2)<br />

in this case, a positive scalar. To make this program solvable by convex<br />

iteration, as explained in Example 4.4.1.2.2 and other previous examples, we<br />

move the rank constraint to the objective<br />

minimize<br />

X , Y , Z , x , y<br />

subject to (x, y) ∈ C<br />

⎡<br />

f(x, y) + ‖X − Y ‖ F + 〈G, W 〉<br />

G = ⎣<br />

tr(X) = 1<br />

δ(Z) ≽ 0<br />

X Z x<br />

Z Y y<br />

x T y T 1<br />

⎤<br />

⎦≽ 0<br />

(727)<br />

by introducing a direction matrix W found from (1581a):<br />

minimize<br />

W ∈ S 2N+1 〈G ⋆ , W 〉<br />

subject to 0 ≼ W ≼ I<br />

trW = 2N<br />

(728)<br />

This semidefinite program has an optimal solution that is known in<br />

closed form. Iteration (727) (728) terminates when rankG = 1 and linear<br />

regularization 〈G, W 〉 vanishes to within some numerical tolerance in (727);<br />

typically, in two iterations. If function f competes too much with the<br />

regularization, positively weighting each regularization term will become<br />

required. At convergence, problem (727) becomes a convex equivalent to<br />

the original nonconvex problem (724).<br />

<br />

4.6.0.0.8 Example. fast max cut. [96]<br />

Let Γ be an n-node graph, and let the arcs (i , j) of the graph be<br />

associated with [ ] weights a ij . The problem is to find a cut of the<br />

largest possible weight, i.e., to partition the set of nodes into two<br />

parts S, S ′ in such a way that the total weight of all arcs linking<br />

S and S ′ (i.e., with one incident node in S and the other one<br />

in S ′ [Figure 85]) is as large as possible. [31,4.3.3]

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

Saved successfully!

Ooh no, something went wrong!