svpChange

svpChange implements Smallest Valid Partitioning (SVP) for multiple change-point detection in univariate time series. SVP combines dynamic programming with a local validity constraint: a segment is retained only when its selected validity statistic stays below a user-defined threshold.

The computational validity tests are implemented in C++ through Rcpp.

Installation

install.packages("remotes")
# remotes::install_github("vrunge/svpChange")
install.packages("svpChange", repos = c('https://vrunge.r-universe.dev', 'https://cloud.r-project.org'))

For local development:

remotes::install_local("path/to/svpChange")

Basic use

library(svpChange)
set.seed(1)
n <- 120
y <- ts_generator(
  chpts = c(40, 80, 120), parameters = c(0, 2, -1),
  sd_noise = 1, type = "gauss"
)

fit <- SVP(y, gamma = 1.5 * log(length(y)), test = "gaussian_mean")
fit$changepoints

changepoints contains the inclusive end index of every segment, including the final observation. R contains the dynamic-programming cost, segment count, and previous boundary for each time point.

Built-in validity tests

SVP() accepts these values of test:

Test	Use
gaussian_mean	Gaussian FOCUS mean-change test
gamma_rate	Change in rate for positive Gamma observations
gaussian_variance	Variance changes through squared observations
quantile, quantileExact	Quantile-based robust tests
varCost	Robust variance cost test
WilcoxonCost	Wilcoxon rank-based test
MedianMoodCost	Median-Mood rank-based test
AR1	Exact fixed-rho AR(1) mean-change test
AR1Profile	AR(1) test with profiled innovation variance
AR1Focus	Faster innovation-based AR(1) approximation

For quantile tests:

SVP(y, gamma = 10, test = "quantile", quantile = 0.05)

For AR(1) data:

fit_ar1 <- SVP(y, gamma = 10, test = "AR1", rho = 0.7, sigma2 = 1)

If rho is omitted, SVP() estimates it robustly using AR1_rho().

Pruning and multiscale thresholds

The subtests argument controls the candidate set:

SVP(y, gamma = 10, test = "gaussian_mean",
    subtests = "both")

The values are "none", "right", and "both". Use "none" for arbitrary validity rules unless the pruning assumptions have been established for the selected test.

User-defined validity tests

Use svp0() when the validity rule is an R function:

my_test <- function(segment, gamma) max(segment) - min(segment) <= gamma
fit <- svp0(y, gamma = 5, test = my_test)

The package also provides valid_FOCUS(), valid_AR1(), valid_SSE(), valid_RANGE(), valid_RANGE_SLACK(), valid_QUANTILE(), valid_SCALE(), and valid_OP() for svp0().

Other algorithms and utilities

OP(), PELT(), and SN() are available for algorithm comparisons. AR1_rho() and AR1_single_change() provide AR(1) diagnostics, and ts_generator() generates simulation signals.

Simulations

The simulations/ directory contains paper-style experiments:

power_gaussian/  Gaussian power studies
power_ar1/       AR(1) power studies
power_robust/    Heavy-tailed power studies
time_gaussian/   Gaussian runtime studies
time_ar1/        AR(1) runtime studies
time_robust/     Robust runtime studies
other_simus/     Supporting, historical, and application studies

Each primary folder contains a README.md. Large Monte Carlo studies are opt-in; see simulations/README.md for the run convention.

Testing

From the package root:

devtools::test()

The tests cover Gaussian, robust, quantile, AR(1), pruning, and algorithm-equivalence behavior.

License

GPL-3. See DESCRIPTION for authorship and package metadata.