iss {Libra}

ISS solver for linear model with lasso penalty

Description

Solver for the entire solution path of coefficients for ISS.

Usage

iss(X, y, intercept = TRUE, normalize = TRUE, nvar = min(dim(X)))

Arguments

X

An n-by-p matrix of predictors

y

Response Variable

intercept

if TRUE, an intercept is included in the model (and not penalized), otherwise no intercept is included. Default is TRUE.

normalize

if normalize, each variable is scaled to have L2 norm square-root n. Default is TRUE.

nvar

Maximal number of variables allowed in the model.

Details

The ISS solver computes the whole regularization path for lasso-penalty for linear model. It gives the piecewise constant solution path for Bregman Inverse Scale Space Differential Inclusion. It is the asymptotic limit of LB method with kaapa goes to infinity and alpha goes to zero.

Value

An "LB" class object is returned. The list contains the call, the family, the path, the intercept term a0 and value for alpha, kappa, iter, and meanvalue, scale factor of X, meanx and normx.

Author(s)

Feng Ruan, Jiechao Xiong and Yuan Yao

References

Ohser, Ruan, Xiong, Yao and Yin, Sparse Recovery via Differential Inclusions, https://arxiv.org/abs/1406.7728

Examples

#Examples in the reference paper
library(MASS)
library(lars)
library(MASS)
library(lars)
n = 80;p = 100;k = 30;sigma = 1
Sigma = 1/(3*p)*matrix(rep(1,p^2),p,p)
diag(Sigma) = 1
A = mvrnorm(n, rep(0, p), Sigma)
u_ref = rep(0,p)
supp_ref = 1:k
u_ref[supp_ref] = rnorm(k)
u_ref[supp_ref] = u_ref[supp_ref]+sign(u_ref[supp_ref])
b = as.vector(A%*%u_ref + sigma*rnorm(n))
lasso = lars(A,b,normalize=FALSE,intercept=FALSE,max.steps=100)
par(mfrow=c(3,2))
matplot(n/lasso$lambda, lasso$beta[1:100,], xlab = bquote(n/lambda), 
        ylab = "Coefficients", xlim=c(0,3),ylim=c(range(lasso$beta)),type='l', main="Lasso")
object = iss(A,b,intercept=FALSE,normalize=FALSE)
plot(object,xlim=c(0,3),main=bquote("ISS"))
kappa_list = c(4,16,64,256)
alpha_list = 1/10/kappa_list
for (i in 1:4){
  object <- lb(A,b,kappa_list[i],alpha_list[i],family="gaussian",group=FALSE,
               trate=20,intercept=FALSE,normalize=FALSE)
  plot(object,xlim=c(0,3),main=bquote(paste("LB ",kappa,"=",.(kappa_list[i]))))
}