Skip to contents

High-dimensional covariance changepoints

Source: R/wrap-highdim2.R
hdcov_wrapper.Rd

Wraps changepoints::BS.cov() (Wang, Yu and Rinaldo): binary segmentation on the sample covariance operator, which detects a change in the dependence structure of a multivariate series even when every marginal mean and variance is unchanged. No other engine in the package can see that.

Usage

hdcov_wrapper(
  x,
  threshold = NULL,
  alpha = 0.05,
  n_perm = 20,
  delta = NULL,
  seed = NULL
)

Arguments

x

A numeric matrix or data frame, rows as time points.

threshold

Detection threshold on the CUSUM statistic. When NULL (the default) it is calibrated by permutation: the time order is shuffled n_perm times, which destroys any changepoint while preserving the marginal distributions, and the threshold is the 1 - alpha quantile of the largest statistic seen.

alpha

Family-wise level for the permutation threshold. Defaults to 0.05.

n_perm

Permutations used to calibrate the threshold. Defaults to 20; raise it for a sharper threshold at proportional cost.

delta

Minimum spacing between changepoints. Defaults to max(10, floor(n / 20)).

seed

Optional seed (the permutation calibration is random).

Value

A ggcpt object with change_in = "covariance"; the changepoints tibble carries the CUSUM statistic in cusum.

References

Wang D, Yu Y, Rinaldo A (2021). “Optimal covariance change point localization in high dimensions.” Bernoulli, 27(1), 554–575. doi:10.3150/20-BEJ1249 .

See also

Other changepoint engines: bcp_wrapper(), beast_wrapper(), bfast_wrapper(), binsegrcpp_wrapper(), bocpd_wrapper(), cpm_wrapper(), cpop_wrapper(), cpt_wrapper(), decafs_wrapper(), ecp_wrapper(), envcpt_wrapper(), esac_wrapper(), fabisearch_wrapper(), fastcpd_wrapper(), fcov_wrapper(), fmean_wrapper(), fpop_wrapper(), geomcp_wrapper(), hdreg_wrapper(), idetect_wrapper(), inspect_wrapper(), kcp_wrapper(), kwc_wrapper(), mcp_wrapper(), mosum_wrapper(), network_wrapper(), not_wrapper(), npmojo_wrapper(), nsp_wrapper(), ocd_wrapper(), pilliat_wrapper(), segmented_wrapper(), smuce_wrapper(), sn_wrapper(), strucchange_wrapper(), taylor_wrapper(), tguh_wrapper(), trend_wrapper(), var_wrapper(), wbs2_wrapper(), wbs_wrapper(), wbsts_wrapper()

Examples

set.seed(2026)
p <- 5
A <- matrix(rnorm(100 * p), ncol = p)
B <- matrix(rnorm(100 * p), ncol = p)
B[, 2] <- B[, 1] + 0.2 * B[, 2]        # correlation appears
hdcov_wrapper(rbind(A, B), n_perm = 20, alpha = 0.05, seed = 1)
#> ggcpt (changepoint detection result)
#>   Method:         hdcov
#>   Change in:       covariance 
#>   Changepoints found: 1 
#>   CP convention:   left 
#>   Penalty:         threshold = 7.666 
#>   Series length:   200 
#> 
#> Changepoints:
#> # A tibble: 1 × 3
#>      cp cp_value cusum
#>   <int>    <dbl> <dbl>
#> 1    93   -0.475  8.22