local.refine.multi.nonpar.L2.epan: Local refinement for multivariate nonparametric change points...

In HaotianXu/changepoints: A Collection of Change-Point Detection Methods

Local refinement for multivariate nonparametric change points localisation based on L2 distance.

Description

Perform local refinement for multivariate nonparametric change points localisation based on L2 distance.

Usage

local.refine.multi.nonpar.L2.epan(
  cpt_init,
  Y,
  kappa_hat,
  r = 2,
  w = 0.9,
  c_kappa = 10
)

Arguments

cpt_init	An integer vector of initial change points estimation (sorted in strictly increasing order).
Y	A numeric matrix of observations with with horizontal axis being time, and vertical axis being dimension.
kappa_hat	A numeric vector of jump sizes estimator.
r	An integer scalar of smoothness parameter of the underlying Holder space.
w	A numeric scalar in (0,1) representing the weight for interval truncation.
c_kappa	A numeric scalar to be multiplied by kappa estimator as the bandwidth in the local refinment.

Value

A vector of locally refined change points estimation.

Author(s)

Haotian Xu

Examples

n = 150
v = c(floor(n/3), 2*floor(n/3)) # location of change points
r = 2
p = 6
Y = matrix(0, p, n) # matrix for data
mu0 = rep(0, p) # mean of the data
mu1 = rep(0, p)
mu1[1:floor(p/2)] = 2
Sigma0 = diag(p) #Covariance matrices of the data
Sigma1 = diag(p)
# Generate data
for(t in 1:n){
  if(t <= v[1] || t > v[2]){
     Y[,t] = MASS::mvrnorm(n = 1, mu0, Sigma0)
  }
  if(t > v[1] && t <= v[2]){
     Y[,t] = MASS::mvrnorm(n = 1, mu1, Sigma1)
  }
}## close for generate data
M = 100
intervals = WBS.intervals(M = M, lower = 1, upper = ncol(Y)) #Random intervals
h = 2*(1/n)^{1/(2*r+p)} # bandwith
temp = WBS.multi.nonpar.L2(Y, 1, ncol(Y), intervals$Alpha, intervals$Beta, h, delta = 15)
cpt_init = tuneBSmultinonpar(temp, Y)
kappa_hat = kappa.multi.nonpar.L2(cpt_init, Y, h_kappa = 0.01)
local.refine.multi.nonpar.L2(cpt_init, Y, kappa_hat, r = 2, w = 0.9, c_kappa = 2)

HaotianXu/changepoints documentation built on Oct. 11, 2023, 12:48 p.m.