Learn R Programming

ICtest (version 0.3-8)

PCAaugHD: High-Dimensional Augmentation Estimate for PCA

Description

Estimates the number of spikes in a spiked covariance model using the high-dimensional predictor augmentation (HDPA) estimator of Radojicic and Virta (2025). In contrast to PCAaug the estimate stays consistent when the dimension and the number of augmented components grow proportionally with the sample size.

Usage

PCAaugHD(X, naug = NULL, nrep = 1, sigma2 = NULL, K = NULL)

Value

A list of class ladle containing:

method

the string HDPA.

k

the estimated value of k.

fn

the rescaled norms \(h_j\), \(j = 1, ..., K+1\).

phin

the debiased eigenvalues \(f(l_j)\), \(j = 1, ..., K+1\).

gn

the main criterion, the differences \(h_{j+1} - h_j\), \(j = 1, ..., K\). k is the value where gn takes its minimum.

lambda

the eigenvalues of the covariance matrix.

W

the transformation matrix to the principal components.

S

data matrix with the centered principal components.

MU

the location of the data which was substracted before calculating the principal components.

data.name

the name of the data for which the augmented estimate was computed.

sigma2

the value used as noise variance.

index.start

the index of the first element of the criteria, used by ladleplot.

Arguments

X

numeric data matrix with n rows and p columns. p can be larger than n.

naug

number of components to be augmented. If NULL, max(p, min(ceiling(n/2), 1000)) is used. See details.

nrep

number of repetitions for the augmentation procedure.

sigma2

the noise variance. If NULL it is estimated from the data, see details.

K

the largest value of k considered. If NULL, min(20, min(p, n-1) - 1) is used.

Author

Klaus Nordhausen

Details

The model assumes that the eigenvalues of the covariance matrix are of the form \(\lambda_1 \geq ... \geq \lambda_{k} > \lambda_{k+1} = ... = \lambda_p = \sigma^2\) and the goal is to estimate the value of k, where \(1 \leq k \leq K\).

Let \(\gamma = p/n\) and \(\rho = r/n\), where \(r\) is the number of augmented components. The augmented estimator of Luo and Li (2021) (see PCAaug) is not consistent when \(\gamma\) or \(\rho\) is not close to zero. Radojicic and Virta (2025) modify it as follows:

  • The eigenvalues \(l_j\) of the covariance matrix of the original data are debiased using \(f(l) = \{l - \sigma^2(1+\gamma)\}/2 + [\{l - \sigma^2(1+\gamma)\}^2 - 4 \sigma^4 \gamma]_+^{1/2}/2\).

  • \(r\) Gaussian components with variance \(\sigma^2\) are added to the data and \(\|u_j\|^2\) denotes the squared norm of the augmented part of the jth eigenvector of the covariance matrix of the augmented data.

  • The rescaled norms \(h_j = f(l_j)\{f(l_j) + (\gamma + \rho)\sigma^2\}/\{f(l_j) + \sigma^2\} \|u_j\|^2\) are asymptotically constant both for \(j \leq k\) and for \(j > k\), and the jump between the two levels is negative.

  • The estimate of k is the value \(j = 1, ..., K\) which minimizes \(h_{j+1} - h_j\).

If sigma2 = NULL, the noise variance is estimated by the median of the non-zero eigenvalues of the covariance matrix, divided by the median of the Marchenko-Pastur distribution with ratio \(\gamma\) if \(\gamma \leq 1\), and by \(\gamma\) times the median of the Marchenko-Pastur distribution with ratio \(1/\gamma\) otherwise. This estimate is biased upwards when k is not small compared to min(p, n)/2. In that case it is better to provide sigma2.

The jump at k grows with \(\rho\), so a large number of augmented components is preferable. The computing time grows however with min(n, p + naug)^3, which is the reason for the cap in the default value of naug. If nrep > 1, the norms \(\|u_j\|^2\) are averaged over the repetitions.

The theory assumes Gaussian data and that the spikes satisfy \(\lambda_k - \sigma^2 > \sigma^2(\gamma + \rho)^{1/2}\). Radojicic and Virta (2025) note that a sample size of at least 500 is needed for reliable estimates.

The criterion can be visualized using ladleplot. Here the index starts at 1 and the element fn contains \(h_j\) and phin the debiased eigenvalues.

References

Luo, W. and Li, B. (2021), On Order Determination by Predictor Augmentation, Biometrika, 108, 557--574. <doi:10.1093/biomet/asaa077>

Radojicic, U. and Virta, J. (2025), Dimension Estimation in a Spiked Covariance Model Using High-Dimensional Data Augmentation, Biometrika, 112, asaf052. <doi:10.1093/biomet/asaf052>

See Also

PCAaug, ladleplot, PCAladle

Examples

Run this code
n <- 500
p <- 100
ev <- c(4, 3, 2, rep(0, p - 3)) + 1
X <- matrix(rnorm(n * p), ncol = p) %*% diag(sqrt(ev))

test <- PCAaugHD(X)
test
summary(test)
ladleplot(test)
ladleplot(test, crit = "fn")

Run the code above in your browser using DataLab