this is for holding javascript data
adam greenberg edited methods.tex
about 10 years ago
Commit id: 1068fa6d34c25ee0bb0f7819ace1b6222521e0e5
deletions | additions
diff --git a/methods.tex b/methods.tex
index 9eba31d..e7a9b3d 100644
--- a/methods.tex
+++ b/methods.tex
...
Note that since $A'$ is upper triangular, its inverse can be easily calculated.
\subsection{Additions to SRIF}
I have implemented 2 main additions to the standard SRIF. An inherent downside to the SRIF is the number of operations necessary to generate the Householder matrix $H$ grows like
$n*m^2$. $m^2(n-m)$. My implementation of the SRIF runs the matrix triangularization simultaneously on multiple cores, which results in a significant speed-up. This can be done in a thread-safe manner because each Householder operation occurs on a column-by-column basis. \par
The second addition I have made to the standard SRIF is the inclusion of a secondary $\chi^2$ minimization for the scaling of $\vec{\delta{x}}$. So
\[Q'' = Q(\vec{x} + \alpha \vec{\delta x})\] is minimized. This minimization is done with a grid search over 6 decades. The additional minimization adds a trivial additional computation cost to the overall minimization of $\chi^2$, but allows for faster convergence, and the possibility of skipping over local minima in the $\chi^2$-space.