README.md

In adallak/SCPackage: Learning Smooth Cholesky Factors of Large Covariance Matrices

SCPackage

Introduction

The SC is a package that estimates a Cholesky factor of inverse covariance matrix via penalized maximum likelihood under local stationary assumption. In particular, given a n-by-p data matrix \eqn{X} , with each row an observation of a p dimensional random vector \eqn{X \sim N(0, \Omega^{-1}) = (L^T L)^{-1})}, this package implements a penalized likelihood-based approach of estimating \eqn{L} by smoothing its subdiagonals. This document serves as an introduction of using the package.

The main function is smoothchol, which takes a sample covariance matrix of the observations and returns the estimate of \eqn{L}.

Installation

To install the latest version from Github, use

library(devtools)
devtools::install_github("adallak/SCPackage")

Usage

The package contains function generateL for generating true standard and modified Cholesky factor $L$. It takes as an input number of variables and number of bands and returns the Cholesky Factor.

library(SC)
set.seed(12)
p <- 50
L_true <- generateL(p = p, case = "b")$L

Having true Cholesky factor, we can then generate a (n x p)data matrix X with each row a random sample drawn independently from a Gaussian distribution of mean zero and covariance \eqn{\Sigma = (L^T L)^{-1}}. We use function sample_gen from the package varband to generate the data.

library(varband)
n = 100
# random sample
X <- sample_gen(L = L_true, n = n)
# sample covariance matrix
S <- crossprod(scale(X, center = TRUE, scale = FALSE)) / n

Estimating L with a fixed tuning parameter

The scfunction takes the following parameters:

• S - sample covariance matrix
• lambda1 - controls the sparsity level
• lambda_2 - controls the smoothness level
• max_iter - Number of iteration for block coordinate algorithm
• init.x - Initial vectorized Cholesky factor \eqn{L}. If NULL, the default is \eqn{vec(\sqrt{diag(S)})}.
• type - type of the smoothing penalty
• band - if specified, algorithm iterates only over specifed number of subdiagonals and forces the rest of entries to be equal to zero.
• ABSTOL - Tolerance for algorithm convergence.

L_fused = smoothchol(S, lambda1 = 0, lambda2 = 0.2, type = "fused")$L
L_l1trend = smoothchol(S, lambda1 = 0, lambda2 = 0.2, type = "l1trend")$L
L_HP = smoothchol(S, lambda1 = 0, lambda2 = 0.2, type = "HP")$L

adallak/SCPackage documentation built on Feb. 7, 2022, 1:58 a.m.