next up previous
Next: Convergence of secant methods Up: A secant update for Previous: Note:

Proof:

The proof is a direct computation:

\begin{eqnarray*}&&\left(A+UV^T\right)\left(A^{-1}-A^{-1}U\left(I+V^TA^{-1}U\rig...
...TA^{-1}U\right)
\left(I+V^TA^{-1}U\right)^{-1}V^TA^{-1}\\
&=&I.
\end{eqnarray*}


QED

I now apply the Sherman-Morrison-Woodbury formula to the matrix Jdefined by (9) by defining

\begin{displaymath}U=\left[\delta y\vert\gamma Hs\right],\ V=\left[L^Ts\vert L^Ts\right],
\end{displaymath}

where

\begin{displaymath}\delta=\frac{1}{\alpha s\cdot Hs}>0,\ \gamma=-\frac{1}{s\cdot Hs}<0.
\end{displaymath}

Then J=L+UVT, and, assuming L is nonsingular, J is nonsingular provided

I+VTL-1U

is nonsingular. But

\begin{displaymath}I+V^TL^{-1}U=\left[\begin{array}{cc}
1+\delta s\cdot y&\gamma...
...}&-1\\
\frac{s\cdot y}{\alpha s\cdot Hs}&0\end{array}\right],
\end{displaymath}

and so

\begin{displaymath}\det\left(I+V^TL^{-1}U\right)=\frac{s\cdot y}{\alpha s\cdot Hs}>0.
\end{displaymath}

Therefore I+VTL-1U is nonsingular and hence so is J.



Mark S. Gockenbach
2003-02-17