DCIi1

I1 mean-change testing and descriptive DCI candidate detection, based on Waychal, Laha and Sinha (2026), arXiv:2610.23476.

Installation

install.packages("DCIi1_0.2.0.tar.gz", repos = NULL, type = "source")
library(DCIi1)

I1 test

Reference p-values require 100 observations, w1=22 and independent Gaussian noise. The input checks enforce length, window and mode; the user assesses the noise assumption. Mean-shift alternatives are allowed.

set.seed(12)
y <- rnorm(100) + rep(c(0, 3), each = 50)
result <- i1_test(y)
result
attr(result, "diagnostics")

The calibration is an empirical density approximation. In 10000 independent Gaussian null validation series, rejection rates were 0.86%, 4.15% and 8.57% at nominal levels 1%, 5% and 10%. This is finite-sample evidence, not a claim of exact size or validity under other distributions.

For other lengths, windows or WD scaling, use statistic-only analysis:

set.seed(13)
long_series <- rnorm(300)
i1_test(long_series, calibration = "none")  # pvalue is NA
i1_stt(long_series)
i1_test(y, archeck = 1, calibration = "none")

DCI candidate locations

DCI flags and refines candidate mean changes. It does not assign p-values to data-selected segments or label candidates statistically significant.

set.seed(21)
y_multi <- rnorm(160) + rep(c(0, 4, -3, 1), c(50, 15, 15, 80))
candidates <- dcii1(y_multi)
locations <- change_points(candidates)
plot(y_multi, type = "l", xlab = "Observation", ylab = "Value")
abline(v = locations, col = "firebrick", lty = 2)
covering_metric_unwtd(160, c(50, 65, 80), locations)

All DCI pvalue entries are NA. Candidate locations can occur under a constant mean. change_points() sorts and deduplicates candidates without testing them. Locations are the last observation before a boundary. Magnitudes are signed standardized contrasts, not original-unit shifts. For I1, multiply cpmag by the diagnostics sigma to recover data units.

Practical scope

• Numeric finite univariate data; no implicit missing-value deletion.
• Default full window 22 means up to 11 observations on each side.
• The WD denominator is -2 * sqrt(2) * (1-rho) * qnorm(.25), as in Appendix B.1.
• Zero IQR, undefined WD estimates and unsupported inferential settings error explicitly.
• No asymptotic p-values, multiplicity adjustment, or general variance-change test.
• No external data downloads, workspace loading or compiled code are needed.

See ?i1_test, ?dcii1, vignette("i1-test", package="DCIi1") and vignette("dci-algorithm", package="DCIi1").

Authors and licence

Nachiketas Waychal, Arnab Kumar Laha and Ankur Sinha are the software authors and copyright holders. Maintainer: Nachiketas Waychal nachiketas.waychal@yahoo.com. Distributed under GPL-3. Affiliations, attribution and AI-assisted development acknowledgement are in inst/NOTICE. Cite the method with citation("DCIi1").