Newton-Raphson

The essential feature of this algorithm for solving tex2html_wrap_inline175 , is that given a guess for the solution, say tex2html_wrap_inline177 , we get an update using the rule

displaymath179

Try as an example to use it to find the value of x such that tex2html_wrap_inline183 .

That is tex2html_wrap_inline185 and tex2html_wrap_inline187 , so that the update is

displaymath189

Start of with a guess, say tex2html_wrap_inline191 then

displaymath193

displaymath195

displaymath197

which can be further iterated, for example, tex2html_wrap_inline199 , tex2html_wrap_inline201 .

The real answer is of course tex2html_wrap_inline203 , so we essentially get there after just 5 iterations.

Where does Newton-Raphson come from? Either a picture, or as a first order Taylor Series approximation. First order Taylor series approximations are often referred to as a linearization of the problem.

The matrix version for the Newton-Raphson comes from thinking about approximating the function using a hyper-plane, and gives an update

displaymath205

where x and g are now vectors and therefore g' is a square symmetric matrix.

Matrix inversion should be avoided wherever possible, so we re-express the update by multiplying through by g' to get

displaymath215

Note that this is now a linear system of equations of the form,

displaymath217

where A is a tex2html_wrap_inline143 matrix, tex2html_wrap_inline223 is a p-vector as is tex2html_wrap_inline227 .

If we can solve this equation for y then clearly tex2html_wrap_inline231 .

This is how we will really do it, and we can use the Q-R decompostion to solve this system of linear equations.


next up previous
Next: Other topics Up: Class 5 Stat 540 Previous: A non-linear least squares

Richard Waterman
Fri Feb 12 06:43:38 EST 1999