next up previous
Next: Bibliography Up: Preconditioned conjugate gradient method Previous: Appendix A

Appendix B

To perform a line minimization from a point $X^{(k)}$ along a certain direction $D^{(k)}$, we wish to find $\lambda_{\mathrm{opt}}$, the optimum value of $\lambda$ which minimizes
\begin{displaymath}
f(\lambda) = \sum_{m=1}^{M}
\frac{ (X^{(k)}_m + \lambda D^{(...
...}_m + \lambda D^{(k)}_m)^T S (X^{(k)}_m + \lambda D^{(k)}_m)}.
\end{displaymath} (27)

This may be achieved in several ways. First, by calculating the derivative of $f$ at $\lambda = 0$, $\left. \frac{\d f}{\d\lambda}
\right\vert _{\lambda = 0}$, taking a trial step $\lambda_{\mathrm{t}}$ to evaluate $f_{\mathrm{t}} = f(\lambda_{\mathrm{t}})$ and making a parabolic fit to determine $\lambda_{\mathrm{opt}}$.

Alternatively, since

\begin{displaymath}
\frac{\d f}{\d\lambda} = \sum_{m=1}^{M} \frac{2(a_m + \lambda b_m +
\lambda^2 c_m)}{(1+ \lambda^2 (D_m)^T S D_m)^2},
\end{displaymath} (28)

where
$\displaystyle a_m$ $\textstyle =$ $\displaystyle \left[\left(X^{(k)}_m\right)^T S X^{(k)}_m\right]
\left[\left(D^{(k)}_m\right)^T H X^{(k)}_m\right] -$  
    $\displaystyle \qquad\qquad \left[\left(X^{(k)}_m\right)^T H X^{(k)}_m\right]
\left[\left(D^{(k)}_m\right)^T SX^{(k)}_m\right],$ (29)
$\displaystyle b_m$ $\textstyle =$ $\displaystyle \left[\left(X^{(k)}_m \right)^T SX^{(k)}_m\right]
\left[\left(D^{(k)}_m \right)^T HD^{(k)}_m\right] -$  
    $\displaystyle \qquad\qquad \left[\left(X^{(k)}_m \right)^T H X^{(k)}_m \right]
\left[\left(D^{(k)}_m \right)^T SD^{(k)}_m \right],$ (30)
$\displaystyle c_m$ $\textstyle =$ $\displaystyle \left[\left(X^{(k)}_m \right)^T SD^{(k)}_m \right]
\left[\left(D^{(k)}_m \right)^T H D^{(k)}_m \right] -$  
    $\displaystyle \qquad\qquad \left[\left(X^{(k)}_m \right)^T H D^{(k)}_m \right]
\left[\left(D^{(k)}_m \right)^T S D^{(k)}_m \right],$ (31)

we find $\lambda_{\mathrm{opt}}$ as one of the roots of the quadratic equation
\begin{displaymath}
a + b \lambda + c \lambda^2 = 0,
\end{displaymath} (32)

where $a = {\displaystyle{\sum_m} a_m}$ etc.
next up previous
Next: Bibliography Up: Preconditioned conjugate gradient method Previous: Appendix A
Peter Haynes