next up previous
Next: Long-step path-following method Up: Numerical realization and results Previous: Numerical realization and results

Linear system

Here, we will discuss some computational issues in second order cone programming, especially the linear system (1-6). Denote by $ [r_1,r_2,r_3,r_4,r_5]^T$ the right hand side of (1-6). The augmented system following from (1-6) is

$\displaystyle \begin{pmatrix}-(\bar{X}T(\Theta W)^{-1})^{-1}\bar{S}T\Theta
W-\f...
...matrix}=\begin{pmatrix}r_2-\Theta
WT\bar{X}^{-1}r_4+mc\cr r_1+mb\end{pmatrix},
$

where $ m=\frac{\tau}{\kappa}(r_3+r_5/\tau).$ We use the NT direction, i.e., $ \bar{X}=\bar{S}$, apply the Sherman-Morrison-Woodury formula to the coefficient matrix, and denote

\begin{displaymath}\begin{array}{rcl} p=\begin{pmatrix}p_1 \cr p_2\end{pmatrix}&...
...x}r2-\Theta WT\bar{X}^{-1}r_4 \cr r_1\end{pmatrix}. \end{array}\end{displaymath} (4-1)

Hence,

$\displaystyle \begin{pmatrix}\triangle x\cr\triangle y\end{pmatrix}=\begin{pmatrix}q_1\cr q_2\end{pmatrix}+h\begin{pmatrix}p_1\cr p_2\end{pmatrix},$    where $\displaystyle h=\frac{m-\frac{\tau}{\kappa}\begin{pmatrix}-c^T & b^T\end{pmatrix}q}{1+\frac{\tau}{\kappa}\begin{pmatrix}-c^T& b^T\end{pmatrix}p}.$ (4-2)

The next lemma follows immediately.

Lemma 4.1   $ h=\triangle \tau .$

Proof.

\begin{displaymath}
\begin{array}{rcl}
\triangle\tau -h&=&\frac{\tau}{\kappa}(r_...
...a}\begin{pmatrix}-c^T
& b^T\end{pmatrix}p)\cr &=&0.
\end{array}\end{displaymath}

Moreover, the normal equations are

\begin{displaymath}
\begin{array}{rcl}
A(\Theta W)^{-2}A^Tp_2&=&b+A(\Theta W)^{-...
...Theta W)^{-2}(r_2-\Theta WT\bar{X}^{-1}r_4-A^Tq_2).
\end{array}\end{displaymath}

Subsequently, the search direction can be computed as follows:

\begin{displaymath}
\begin{array}{rcl}
\triangle\tau&=&h,\cr \triangle x&=& q_1+...
...kappa&=&-\frac{\kappa}{\tau}\triangle\tau+r_5/\tau.
\end{array}\end{displaymath}

Furthermore, one can rewrite $ \triangle s$ as

$\displaystyle \triangle s=r_2-A^Tq_2+\triangle\tau(c-A^Tp_2).$

Two vectors $ r_2-A^Tq_2$ and $ c-A^Tp_2$ have been calculated while computing $ p_1$ and $ q_1.$ Because of the definition of $ W^i$, $ (W^i)^{2}$ and $ (W^i)^{-2}$ are easy to obtain:

\begin{displaymath}
\begin{array}{rcl}
(W^i)^{-2}&=&-Q^i-2Q^iw^i(1-2(w^i)^TQ^iw...
...\begin{pmatrix}-w^i_1&(w^i_{2:n_i})^T\end{pmatrix},
\end{array}\end{displaymath}

since $ 1-2(w^i)^TQ^iw^i=-1.$
next up previous
Next: Long-step path-following method Up: Numerical realization and results Previous: Numerical realization and results
Hans D. Mittelmann 2003-09-10