25.11.2015 Views

The MOSEK Python optimizer API manual Version 7.0 (Revision 141)

Optimizer API for Python - Documentation - Mosek

Optimizer API for Python - Documentation - Mosek

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.

11.2. LINEAR OPTIMIZATION <strong>141</strong><br />

Whenever the trial solution satisfies the criterion<br />

yk<br />

AT<br />

∥<br />

( ∣ ∣ (x k ) T s k ∣∣∣<br />

min<br />

(τ k ) 2 , c T x k<br />

τ k − bT y k ) ∣∣∣<br />

τ k<br />

the interior-point <strong>optimizer</strong> is terminated and<br />

x k , s k , τ k , κ k > 0.<br />

∥ ∥∥∥ ∥ Axk τ k − b ∞ ≤ ɛ p (1 + ‖b‖ ∞ ),<br />

∥<br />

τ k + sk ∥∥∥<br />

τ k − c ∞ ≤ ɛ d (1 + ‖c‖ ∞ ), and<br />

(<br />

≤<br />

ɛ g max<br />

1, min(∣ ∣c T x k∣ ∣ , ∣ ∣b T y k∣ )<br />

∣ )<br />

τ k ,<br />

(11.5)<br />

(x k , y k , s k )<br />

τ k<br />

is reported as the primal-dual optimal solution. <strong>The</strong> interpretation of (11.5) is that the <strong>optimizer</strong> is<br />

terminated if<br />

• xk<br />

τ k<br />

•<br />

is approximately primal feasible,<br />

( )<br />

y<br />

k<br />

, sk<br />

τ k τ k<br />

is approximately dual feasible, and<br />

• the duality gap is almost zero.<br />

On the other hand, if the trial solution satisfies<br />

−ɛ i c T x k ><br />

‖c‖ ∞<br />

∥ ∥ Ax<br />

k ∞<br />

max(1, ‖b‖ ∞ )<br />

then the problem is declared dual infeasible and x k is reported as a certificate of dual infeasibility.<br />

<strong>The</strong> motivation for this stopping criterion is as follows: First assume that ∥ ∥ Ax<br />

k ∥ ∥ ∞ = 0 ; then x k is<br />

an exact certificate of dual infeasibility. Next assume that this is not the case, i.e.<br />

and define<br />

∥ Ax<br />

k ∥ ∥ ∞ > 0,<br />

It is easy to verify that<br />

¯x := ɛ i<br />

max(1, ‖b‖ ∞ )<br />

‖Ax k ‖ ∞ ‖c‖ ∞<br />

x k .<br />

‖A¯x‖ ∞ = ɛ i<br />

max(1, ‖b‖ ∞ )<br />

‖c‖ ∞<br />

and − c T ¯x > 1,<br />

which shows ¯x is an approximate certificate of dual infeasibility where ɛ i controls the quality of the<br />

approximation. A smaller value means a better approximation.

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

Saved successfully!

Ooh no, something went wrong!