# example_VAR

library(changepoints)

This is a simple guide for offline change points detection in transition matrix of a VAR1 model.

Function simu.VAR1 simulates a segment of data from a VAR1 model with fixed transition matrix. We prepare data with change points by combining several such segments with different transition matrices.

Dynamic programming and local refinement are implemented for VAR1 change points detection:

1. DP.VAR1: performs dynamic programming for VAR1 change points detection.
• CV.search.DP.VAR1: performs grid search to select tuning parameters (gamma associated with l_0 penalty and lambda associated with l_1 penalty) through cross-validation.
2. local.refine.VAR1: performs local refinement for an initial change points estimation.
• local.refine.CV.VAR1: performs grid search to select the tuning parameter zeta associated with the group lasso penalty through cross-validation.

# Simulate data

## parameters for simulating data
rho = 0.4 #candidate of rho (rho can not be too large so that the time series is stable)
p = 20 # dimension
sigma = 1 # standard deviation of error terms
delta1 = 10 # minimum gap for DP
delta.local = 10 # minimum gap for local.refine
n = 30 # length for each segment
v1 = 2*(seq(1,p,1)%%2) - 1
v2 = -v1
AA = matrix(0, nrow = p, ncol = p-2)
A1 = rho * cbind(v1,v2,AA) # transition matrix for the first segment
A2 = rho * cbind(v2,v1,AA) # transition matrix for the second segment
A3 = A1 # transition matrix for the third segment

set.seed(123)
# generate data
data = simu.VAR1(sigma, p, 2*n+1, A1)
data = cbind(data, simu.VAR1(sigma, p, 2*n, A2, vzero=c(data[,ncol(data)])))
data = cbind(data, simu.VAR1(sigma, p, 2*n, A3, vzero=c(data[,ncol(data)])))
cpt_true = c(2*n, 4*n)

# Perform dynamic programming

gamma_set = c(0.1, 0.5, 1, 2, 5) # a set of tuning parameters for DP
lambda_set = c(0.1, 0.5, 1, 2, 5) # a set of tuning parameters for lasso
DP_result = CV.search.DP.VAR1(data, gamma_set, lambda_set, delta1) # grid search through cross-validation
min_idx = as.vector(arrayInd(which.min(DP_result$test_error), dim(DP_result$test_error)))# select gamma achieves the minimum validation error
cpt_DP_hat = unlist(DP_result$cpt_hat[min_idx[1], min_idx[2]]) # estimated changepoints by DP cpt_DP_hat #> [1] 66 124 Hausdorff.dist(cpt_DP_hat, cpt_true) #> [1] 6 zeta_set = c(1, 1.5, 2) # tuning parameter for group lasso cpt_DPlr_hat = local.refine.CV.VAR1(cpt_DP_hat, data, zeta_set, delta.local) # perform local refinement cpt_DPlr_hat #>$cpt_hat
#> [1]  60 120
#>
#> $zeta #> [1] 1.5 Hausdorff.dist(cpt_DPlr_hat$cpt_hat, cpt_true)
#> [1] 0