Daniel Lakeland wrote:
> Perhaps you can store the value of the
> function, and the gradient of the function at a set of regularly
> spaced grid points (that would be 3 numbers for each grid point, one
> for the function and two for the components of the gradient), and then
> you evaluate the function by finding the closest grid point, and using
> linear extrapolation with the gradient you refine the computation.
This would be the mathematical solution for the problem, using
real numbers.
A fixedpoint solution itself will produce additional errors, that
are missing in such an approximation model. To add them, leads to
the same kind of problem, I'm currently investigating: error
functions like beds of nail.
Using fixedpoint (which means integers to be devided by a factor
at the end and balancing subexpressions to avoid destortions) is
much cheaper in terms of computational power and memory, than
using some float format, if no floatingpoint processor is
available and if the computational errors are tolerable.
It seems to me, that dew point is a good candidate to get it
solved: looks complicated in the original formula, but isn't.