refund (version 0.1-37)

fpca.ssvd: Smoothed FPCA via iterative penalized rank one SVDs.

Description

Implements the algorithm of Huang, Shen, Buja (2008) for finding smooth right singular vectors of a matrix X containing (contaminated) evaluations of functional random variables on a regular, equidistant grid. If the number of smooth SVs to extract is not specified, the function hazards a guess for the appropriate number based on the asymptotically optimal truncation threshold under the assumption of a low rank matrix contaminated with i.i.d. Gaussian noise with unknown variance derived in Donoho, Gavish (2013). Please note that Donoho, Gavish (2013) should be regarded as experimental for functional PCA, and will typically not work well if you have more observations than grid points.

Usage

fpca.ssvd(
  Y = NULL,
  ydata = NULL,
  argvals = NULL,
  npc = NA,
  center = TRUE,
  maxiter = 15,
  tol = 1e-04,
  diffpen = 3,
  gridsearch = TRUE,
  alphagrid = 1.5^(-20:40),
  lower.alpha = 1e-05,
  upper.alpha = 1e+07,
  verbose = FALSE,
  integration = "trapezoidal"
)

Value

an fpca object like that returned from fpca.sc, with entries Yhat, the smoothed trajectories, Y, the observed data, scores, the estimated FPC loadings, mu, the column means of Y (or a vector of zeroes if !center), efunctions, the estimated smooth FPCs (note that these are orthonormal vectors, not evaluations of orthonormal functions if argvals is not equidistant),

evalues, their associated eigenvalues, and npc, the number of smooth components that were extracted.

Arguments

Y

data matrix (rows: observations; columns: grid of eval. points)

ydata

a data frame ydata representing irregularly observed functions. NOT IMPLEMENTED for this method.

argvals

the argument values of the function evaluations in Y, defaults to a equidistant grid from 0 to 1. See Details.

npc

how many smooth SVs to try to extract, if NA (the default) the hard thresholding rule of Donoho, Gavish (2013) is used (see Details, References).

center

center Y so that its column-means are 0? Defaults to TRUE

maxiter

how many iterations of the power algorithm to perform at most (defaults to 15)

tol

convergence tolerance for power algorithm (defaults to 1e-4)

diffpen

difference penalty order controlling the desired smoothness of the right singular vectors, defaults to 3 (i.e., deviations from local quadratic polynomials).

gridsearch

use optimize or a grid search to find GCV-optimal smoothing parameters? defaults to TRUE.

alphagrid

grid of smoothing parameter values for grid search

lower.alpha

lower limit for for smoothing parameter if !gridsearch

upper.alpha

upper limit for smoothing parameter if !gridsearch

verbose

generate graphical summary of progress and diagnostic messages? defaults to FALSE

integration

ignored, see Details.

Author

Fabian Scheipl

Details

Note that fpca.ssvd computes smoothed orthonormal eigenvectors of the supplied function evaluations (and associated scores), not (!) evaluations of the smoothed orthonormal eigenfunctions. The smoothed orthonormal eigenvectors are then rescaled by the length of the domain defined by argvals to have a quadratic integral approximately equal to one (instead of crossproduct equal to one), so they approximate the behavior of smooth eigenfunctions. If argvals is not equidistant, fpca.ssvd will simply return the smoothed eigenvectors without rescaling, with a warning.

References

Huang, J. Z., Shen, H., and Buja, A. (2008). Functional principal components analysis via penalized rank one approximation. Electronic Journal of Statistics, 2, 678-695

Donoho, D.L., and Gavish, M. (2013). The Optimal Hard Threshold for Singular Values is 4/sqrt(3). eprint arXiv:1305.5870. Available from https://arxiv.org/abs/1305.5870.

See Also

fpca.sc and fpca.face for FPCA based on smoothing a covariance estimate; fpca2s for a faster SVD-based approach.

Examples

 ## as in Sec. 6.2 of Huang, Shen, Buja (2008):
 set.seed(2678695)
 n <- 101
 m <- 101
 s1 <- 20
 s2 <- 10
 s <- 4
 t <- seq(-1, 1, l=m)
 v1 <- t + sin(pi*t)
 v2 <- cos(3*pi*t)
 V <- cbind(v1/sqrt(sum(v1^2)), v2/sqrt(sum(v2^2)))
 U <- matrix(rnorm(n*2), n, 2)
 D <- diag(c(s1^2, s2^2))
 eps <- matrix(rnorm(m*n, sd=s), n, m)
 Y <- U%*%D%*%t(V) + eps

smoothSV <- fpca.ssvd(Y, verbose=TRUE)

 layout(t(matrix(1:4, nr=2)))
 clrs <- sapply(rainbow(n), function(c)
           do.call(rgb, as.list(c(col2rgb(c)/255, .1))))
 matplot(V, type="l", lty=1, col=1:2, xlab="",
         main="FPCs: true", bty="n")
 matplot(smoothSV$efunctions, type="l", lty=1, col=1:5, xlab="",
         main="FPCs: estimate", bty="n")
 matplot(1:m, t(U%*%D%*%t(V)), type="l", lty=1, col=clrs, xlab="", ylab="",
         main="true smooth Y", bty="n")
 matplot(1:m, t(smoothSV$Yhat), xlab="", ylab="",
         type="l", lty=1,col=clrs, main="estimated smooth Y", bty="n")