Newton's method in one dimension applied
to the problem of minimizing a function with derivatives and involves approximating the first derivative everywhere in
terms of its value and that of the second derivative at a single evaluated x
coordinate, so that
Equation 1
Then
the x coordinate at which f is extremal is the one for which the first
derivative is zero. The x coordinate
value for which the first derivative of f is zero is easily found from equation
1 as
Equation 2
The
classic problem with this method comes from the fact that the second derivative
at x0 is too small
This
results in overshooting the location of the minimum and in fact frequently
finding a value of f(x) that is worse than the original. After evaluating f at the predicted value we
now have . The best estimate
of is now
Equation
3
This
can be done in terms of a Marquardt parameter such that
Equation
4
Then substituting in equation 2 and redoing the step.
In multi-dimensions, equation 4 is
generalized to find
Equation 5
In
practice at each successful minimization step the smoothers are moved half way
back to 1 on a logarithmic scale. The
ratios of these smoothers are on the order of 1 to 1020.