sparseLowRankiCov: Sparse plus Low Rank inverse covariance

View source: R/SparseLowRankICov.R

sparseLowRankiCovR Documentation

Sparse plus Low Rank inverse covariance

Description

Select an inverse covariance matrix that is a sparse plus low rank decomposition.

Usage

sparseLowRankiCov(data, npn = FALSE, verbose = FALSE, cor = FALSE, ...)

Arguments

data

the n x p data matrix

npn

flag to first fit nonparametric normal transform to the data

verbose

flag to turn on verbose output

cor

flag to use correlation matrix as the input (default: false - uses covariance)

...

arguments to override default algorithm settings (see details)

Details

This is a wrapper function for sparse plus low rank iCov estimations performed by a custom ADMM algorithm.

Therefore, arguments ... should be named. Typically, these are for specifying a penalty parameter, lambda, or the number of penalties to use. By default 10 pentalties are used, ranging logarithmically between lambda.min.ratio*MAX and MAX. Max is the theoretical upper bound on lambda and us max|S|, the maximum absolute value in the data correlation matrix. lambda.min.ratio is 1e-3 by default. Lower values of lambda require more memory/cpu time to compute, and sometimes huge will throw an error.

The argument nlambda determines the number of penalties - somewhere between 10-100 is usually good, depending on how the values of empirical correlation are distributed.#' @export

One of beta (penalty for the nuclear norm) or r (number of ranks) should be supplied or r=2 is chosen by default.


zdk123/SpiecEasi documentation built on Oct. 20, 2023, 6:49 a.m.