next up previous contents
Next: 4.1.5 Finite-Difference Approximations of Up: 4.1 Optimization Methods Previous: 4.1.3 Steepest-descent


4.1.4 Newton Direction

For the Newton direction a quadratic approximation of the target function (4.5) is used. For existing second derivatives of $f(\vec{x})$ the minimum of the target function is found by

\begin{displaymath}
{\mathop{\nabla }\nolimits f(\vec{x}_0)} +
{\mathop{\nabla }\nolimits ^2 f(\vec{x}_0)} \vec{p} = 0
.
\end{displaymath} (4.9)

So the Newton search direction is given by the equation

\begin{displaymath}
{\mathop{\nabla }\nolimits ^2 f(\vec{x}_0)} \vec{p} = - {\mathop{\nabla }\nolimits f(\vec{x}_0)}
\end{displaymath} (4.10)

using the gradient and the Hessian of the target function.

These algorithms need the derivatives of the model function $f(\vec{x})$. If the function $f(\vec{x})$ is given by an analytical formula, the first and second derivatives can be calculated directly. But for a model function formed by results of several simulations, it is impossible to find analytical expressions for the derivatives. For this reason approximation methods have to be used.




R. Plasun