tests/00-auer.R

In PCDimension: Finding the Number of Significant Principal Components

#######################################################
library(PCDimension)
# simulate one unstructured dataset
suppressWarnings( RNGversion("3.5.0") )
set.seed(992500)
NC <- 15
NS <- 200
ranData <- matrix(rnorm(NS*NC, 6), ncol=NC)
# perform PCA and get the set of variances/eigenvectors
spca <- SamplePCA(ranData)
lam <- spca@variances[1:(NC-1)] # remove 0 eigenvalue
rg <- 0:(NC-2)

######################################################
# Test some internal routines

# Theory says that fhat is non-decreasing
fh <- sapply(rg, PCDimension:::fhat, Lambda=lam)
all( diff(fh) >= 0 ) # should be 'TRUE'

plot( rg, fh, type='b', pch=16 )

# Theory says that dhat is non-increasing
thetas <- seq(0, 0.2, by=0.01)
dset <- sapply(thetas, PCDimension:::dhat, Lambda=lam)
all( diff(dset) <= 0 ) # should be 'TRUE'

plot(thetas, dset, type='l')

######################################################
# test external interface on random data

# Auer-Gervini Bayesian method
obj <- AuerGervini(spca)
summary(obj)
agDimension(obj)

plot(obj)

# broken-stick model
bs <- brokenStick(1:NC, NC)
bs0 <- brokenStick(1:(NC-1), (NC-1))

pts <- screeplot(spca, ylim=c(0, 0.2))
lines(pts, bs, type='b', col='blue', lwd=2, pch=16)
lines(pts[-NC], bs0, type='b', col='red', lwd=2, pch=16)

# randomization based model
rndLambdaF(ranData)

rm(spca)
######################################################
# get the previously computed simulated SamplePCA object
data(spca)

vars <- spca@variances
obj <- AuerGervini(vars, dd=dim(spca@scores))
summary(obj)

plot(obj)

f <- makeAgCpmFun("Exponential")
agfuns <- list(twice=agDimTwiceMean,
               km=agDimKmeans, km3=agDimKmeans3, 
               tt=agDimTtest, tt2=agDimTtest2,
               cpt=agDimCPT, cpm=f)
plot(obj, agfuns)

compareAgDimMethods(obj, agfuns)