chebpol (version 2.1-2)

ipol: Create interpolating function.

Description

Create an interpolating function from given values. Several interpolation methods are supported.

Usage

ipol(val, dims = NULL, intervals = NULL, grid = NULL, knots = NULL,
  k = NULL, method = c("chebyshev", "multilinear", "fh", "uniform",
  "general", "polyharmonic", "simplexlinear", "hstalker", "stalker",
  "crbf", "oldstalker"), ...)

Arguments

val

array or function. Function values on a grid, or the function itself. If it is the values, the "dim"-attribute must be appropriately set. If it is a function, it will be evaluated in the grid points.

dims

integer vector. The number of grid points in each dimension. Not needed if val is an array or grid is used.

intervals

list of length 2 numeric vectors. The lower and upper bound in each dimension. Not used if grid is specified.

grid

list. Each element is a vector of ordered grid-points for a dimension. These need not be Chebyshev-knots, nor evenly spaced.

knots

matrix. Each column is a point in an M-dimensional space.

k

numeric. Additional value, used with some methods.

method

character. The interpolation method to use.

...

Further arguments to the function, if is.function(val). And some extra arguments for interpolant creation described in section Details.

Value

A function(x, threads=getOption('chebpol.threads')) defined on a hypercube, an interpolant for the given function. The argument x can be a matrix of column vectors which are evaluated in parallel in a number of threads. The function yields values for arguments outside the hypercube as well, though it will typically be a poor approximation. threads is an integer specifying the number of parallel threads which should be used when evaluating a matrix of column vectors.

Details

ipol is a wrapper around various interpolation methods in package chebpol. Which arguments to specify depends on the method. The interpolation methods are described in vignette("chebpol",package="chebpol").

The method "chebyshev" needs only the number of Chebyshev knots in each dimension. This is either the "dim" attribute of the array val, or the dims argument if val is a function. Also the intervals can be specified if different from [-1, 1].

The method "uniform" is similar to the "chebyshev", but uniformly spaced knots are created. The argument intervals generally goes with dims when something else than standard intervals [-1, 1] are used.

The methods "multilinear", "fh" (Floater-Hormann), "stalker", "hstalker", and "general" needs the argument grid. These are the methods which can use arbitrary Cartesian grids. The stalker spline is described in vignette("stalker",package="chebpol"). The Floater-Hormann method ("fh") also needs the k argument, the degree of the blending polynomials. It defaults to 4.

The method "polyharmonic" needs the arguments knots and k. In addition it can take the logical argument normalize for normalizing the knots to the unit hypercube. The default is NA, which uses normalization if any of the knots are outside the unit hypercube. Also, a logical nowarn is accepted, it is used to suppress a warning in case the system can't be solved exactly and a least squares fallback method is used.

The method "simplexlinear" needs the argument knots. It creates a Delaunay triangulation from the knots, and does linear interpolation in each simplex by weighting the vertex values with the barycentric coordinates.

If knots are required, but the grid argument is given, knots are constructed as t(expand.grid(grid))

The "crbf" is the multilayer compact radial basis function interpolation in ALGLIB (http://www.alglib.net/interpolation/fastrbf.php). It is only available if ALGLIB was available at compile time. It takes the extra arguments "rbase", "layers", and "lambda". These are discussed in the ALGLIB documentation.

There are also some usage examples and more in vignette("chebpol") and vignette('chebusage').

Examples

# NOT RUN {
## evenly spaced grid-points
su <- seq(0,1,length.out=10)
## irregularly spaced grid-points
s <- su^3
## create approximation on the irregularly spaced grid
ml1 <- ipol(exp, grid=list(s), method='multilin')
fh1 <- ipol(exp, grid=list(s), method='fh')
## test it, since exp is convex, the linear approximation lies above
## the exp between the grid points
ml1(su) - exp(su)
fh1(su) - exp(su)

## multi dimensional approximation
f <- function(x) 10/(1+25*mean(x^2))
# a 3-dimensional 10x10x10 grid, first and third coordinate are non-uniform
grid <- list(s, su, sort(1-s))

# make multilinear, Floater-Hormann, Chebyshev and polyharmonic spline.
ml2 <- ipol(f, grid=grid, method='multilin')
fh2 <- ipol(f, grid=grid, method='fh')
hst <- ipol(f, grid=grid, method='hstalker')
ch2 <- ipol(f, dims=c(10,10,10), intervals=list(0:1,0:1,0:1), method='cheb')
knots <- matrix(runif(3*1000),3)
ph2 <- ipol(f, knots=knots, k=2, method='poly')
sl2 <- ipol(f, knots=knots, method='simplexlinear')
# my alglib is a bit slow, so stick to 100 knots
if(havealglib()) crb <- ipol(f, knots=knots[,1:100], method='crbf',
  rbase=2, layers=5, lambda=0) 
# make 7 points in R3 to test them on
m <- matrix(runif(3*7),3)
rbind(true=apply(m,2,f), ml=ml2(m), fh=fh2(m), cheb=ch2(m), poly=ph2(m), sl=sl2(m),hst=hst(m),
crbf=if(havealglib()) crb(m) else NULL )

# }