landmarkreg: Landmark Registration of Functional Observations

Description

It is common to see that among a set of functions certain prominent features such peaks and valleys, called $landmarks$, do not occur at the same times, or other argument values. This is called $phase variation$, and it can be essential to align these features before proceeding with further functional data analyses. This function uses the timings of these features to align or register the curves. The registration involves estimating a nonlinear transformation of the argument continuum for each functional observation. This transformation is called a warping function. It must be strictly increasing and smooth.

Usage

landmarkreg(fdobj, ximarks, x0marks=xmeanmarks,
            WfdPar, monwrd=FALSE, ylambda=1e-10)

Arguments

fdobj

a functional data object containing the curves to be registered.

ximarks

a matrix containing the timings or argument values associated with the landmarks for the observations in fd to be registered. The number of rows N equals the number of observations, and the number of columns NL equals the numbe

x0marks

a vector of length NL of times of landmarks for target curve. If not supplied, the mean of the landmark times in ximarks is used.

WfdPar

a functional parameter object defining the warping functions that transform time in order to register the curves.

monwrd

A logical value: if TRUE, the warping function is estimated using a monotone smoothing methhod; otherwise, a regular smoothing method is used, which is not guaranteed to give strictly monotonic warping functions.

ylambda

Smoothing parameter controlling the smoothness of the registered functions. It can happen with high dimensional bases that local wiggles can appear in the registered curves or their derivatives that are not seen in the unregistered version

Value

a named list of length 2 with components:
fdrega functional data object for the registered curves.
warpfda functional data object for the warping functions.

Details

It is essential that the location of every landmark be clearly defined in each of the curves as well as the template function. If this is not the case, consider using the continuous registration function register.fd. Although requiring that a monotone smoother be used to estimate the warping functions is safer, it adds considerably to the computatation time since monotone smoothing is itself an iterative process. It is usually better to try an initial registration with this feature to see if there are any failures of monotonicity. Moreover, monotonicity failures can usually be cured by increasing the smoothing parameter defining WfdPar. Not much curvature is usually required in the warping functions, so a rather low power basis, usually B-splines, is suitable for defining the functional paramter argument WfdPar. A registration with a few prominent landmarks is often a good preliminary to using the more sophisticated but more lengthy process in register.fd.

Examples

Run this code

#See the analysis for the lip data in the examples.

## setting parameters
library(lattice)
data(landmark.reg.expData) ## containg an simple object called sampleData

eps <- .Machine$double.eps
from <- -1.0187
to <- 9.4551
nb <- 201
nbreaks <- 11
## assign landmarks
landmark <- matrix(c(0.4999, 0.657, 0.8141, 0.5523, 0.5523,
                     3.3279, 3.066, 3.0137, 3.2231, 3.2231),
                   ncol=2)
wbasis <- create.bspline.basis(rangeval=c(from, to),
	                       norder=4, breaks=seq(from, to, len=nbreaks))
WfdPar <- fdPar(wbasis, 1, 1e-4)
## get the density of the data
x <- split(sampleData, factor(sampleData$which))
densY <- sapply(x, function(z){
            r <- range(z[, 1])
            z <- z[, 1]
            z <- z[z>r[1]+eps & z<r[2]-eps]
            density(z, from=from, to=to, n=nb, na.rm=TRUE)$y
        })

argvals <- seq(from, to, len=nb) 
fdobj   <- smooth.basis(argvals, densY, wbasis, 
                   fdnames = c("x", "samples", "density"))$fd

regDens <- landmarkreg(fdobj, landmark, WfdPar=WfdPar, monwrd=TRUE)
warpfdobj <- regDens$warpfd
warpedX <- eval.fd(warpfdobj, argvals)
funs <-  apply(warpedX, 2, approxfun, argvals)
newDat <- list()
for (i in 1:length(funs))
  newDat[[names(funs)[i]]] <- data.frame(data=funs[[i]](x[[i]][,1]),
                                         which=names(funs)[i])
## visualization
lattice::densityplot(~ data | which, sampleData,
                    layout=c(1,5), col="#A6CEE3", n=120,
                     main="Original data set")
dev.new()
lattice::densityplot( ~ data | which, layout=c(1,length(newDat)),
                     do.call(rbind, newDat), col="#A6CEE3", n=120,
                     main="After warping")