[R] multidimensional point.in.polygon??

[Keith Jewell]

Hi,

I seek to identify those points in/outside a multidimensional convex hull 
(geometry::convhulln). Any suggestions?

Background just in case I'm going down a really wrong road:

Given an observed data set with one dependent/observed variable (Y) and 
multiple (3 to 10) independent/design variables (X1, X2, ...) I want to 
increase the number of points by interpolating. I'm using expand.grid(X) to 
generate the X points and locfit::predict.locfit to interpolate the Y 
values. No problem so far.

BUT my observed X data set is very far from rectangular, so the set of 
points produced by expand.grid includes many  "extrapolations" which I'd 
want to remove. I'm aware of the problems of defining extrapolation in 
multiple dimensions and am minded to define it as "outside the convex hull", 
hence my question.

In searching r-project.org I came across a thread in which baptiste auguie 
said "one general way to do this would be to compute the convex hull 
(?chull) of the augmented set of points and test if the point belongs to 
it"; an approach I'd considered generalising to multiple points thus (pseudo 
R code)...
----------------
  baseHull <- convhulln(baseSet)
  augHull <- convhulln(augSet)
  while (augHull != baseHull)
    {   augSet <- augSet [-(augHull !%in% baseHull)]
        augHull <- convhulln(augSet)
    }
--------------------
... but this has that horrible loop including a convhulln!

In the (real, typical) test data set I'm using for development baseSet is 5 
columns by 2637 rows while baseHull has only 42 distinct points. If augHull 
has a similarly simple hull, then each time round the loop will remove only 
a few dozen points; I need to remove many thousands.

And (in my naivete) I think there must be a better way of testing for a 
point inside a polygon, I'd have thought convhulln must need to do that 
often!

Thanks in advance for any pointers,

Keith Jewell