R implementation of the procedure introduced in the paper "Fast and Optimal Inference for Change Points in Piecewise Polynomials via Differencing" available via arXiv.
Given a vector of observations from a one dimensional signal + noise model, where the signal is a piecewise polynomial function of known degree, the procedure returns disjoint intervals which must each contain a change point location uniformly at some significance level specified by the parameter alpha. By default the contaminating noise is assumed to be independently distributed and Gaussian. However, this behaviour can be changed by setting the parameters gaussian_noise and independent_noise appropriately.
The R package can be installed this from this repository using devtools.
devtools::install_github("gaviosha/ChangePointInference")library(ChangePointInference)
set.seed(42)
blocks <- c(rep(0,205),rep(14.64,62),rep(-3.66,41),rep(7.32,164),rep(-7.32,40))
yy <- blocks + rnorm(length(blocks), sd = 7.5)
diffInf_obj <- diffInf(yy, degree = 0, alpha = 0.1)
diffInf_obj |> plot(type = "l", col = "grey")
blocks |> lines(lty = 2, col = "red")
library(ChangePointInference)
set.seed(42)
waves_signal <- c((1:150) * (2**-3), (150:1) * (2**-3), (1:150) * (2**-3), (150:1) * (2**-3))
yy <- waves_signal + 2 * arima.sim(model = list(ar = 0.25), n = length(waves_signal))
diffInf_obj <- diffInf(yy, degree = 1, alpha = 0.1, independent_noise = FALSE)
diffInf_obj |> plot(type = "l", col = "grey")
waves_signal |> lines(lty = 2, lwd = 2, col = "red")