[R] Solving A System of Equations

Thomas W Blackwell tblackw at umich.edu
Wed Apr 9 17:14:20 CEST 2003


Maybe there are other R tools, but what I do, repeatedly, is
to work out first and second derivatives w.r.t. x on paper.
Then I execute one step of Newton's method in R, line by line
at the command line, however many lines it takes.  Then use
up-arrow to repeat those command lines for maybe five or six
iterations of Newton's method.  Quick, clear and cheap.

The point is:  start with a four dimensional grid of starting
values for x,y,z,zz.  This can be a four-dimensional array in
which each variable is initially constant wrt the other three.
In subsequent iterations, particular x,y,z,zz quadruples will
either converge toward a solution or diverge, and you can use
standard tests to distinguish between the two.

After you've done this once or twice by hand, at the command
line, then you'll be ready to code up a single Newton step
as a function, then call it an appropriate (fixed) number of
times inside an outer function.  I wouldn't worry about trying
to be adaptive and economize on computer cycles.  All of that
stuff from the older numerical analysis literature is misplaced
effort, in my view, in this day and age.

Other users may know of much fancier methods.

-  tom blackwell  -  u michigan medical school  -  ann arbor  -

On Tue, 8 Apr 2003, David Tyler wrote:

> I'm trying to solve a system of 3 equations
> as part of a sub-routine in R, ie
> first eqn a/x-b*sqrtx+c=log(1/dx+1/e(sqrtx);
> snd eqn (f*y)/z-g/y-h=-log(2/x+(z/y)/(i*x)
> and third eqn is of the form zz=x/(j-k(z/y)
> where a..k inclusive are constants, x,y,z
> and zz are inputs.
>
> How can this be done in R?



More information about the R-help mailing list