gSeg

Graph-based scans for a single change-point or one changed interval in a sequence of multivariate or object-valued observations.

Default use

library(gSeg)
set.seed(1)
x <- rbind(matrix(rnorm(200), 40), matrix(rnorm(200, 0.5), 40))
fit <- gseg1(x)
fit$k                       # floor(sqrt(nrow(x))) = 8
fit$scanZ$max.type$tauhat
fit$pval.appr$max.type

gseg1() accepts observations in rows; a numeric vector is a one-dimensional sequence. gseg2() uses the same inputs to scan for one changed interval. Both construct a k-MST with k = floor(sqrt(N)) and report the max-type scan (MET) by default. k may be set explicitly. The graph and realized k are returned, so the representation used in an analysis can be recorded.

For non-Euclidean observations, supply pairwise dissimilarities:

fit <- gseg1(dissimilarity = dist(x))

statistics = "g", "w", "o", or "all" selects GET, WET, OET, or all four scans. Analytic p-values are approximations, not exact finite-sample calibration. For exchangeable observations, Monte Carlo permutation p-values can be requested with pval.appr = FALSE, pval.perm = TRUE, B = 999. The permutation test compares scan maxima and uses the plus-one correction with ties included. Arbitrarily serially dependent observations are not made exchangeable by this interface; ordinary permutations are not justified merely by calling these functions. Input validation does not certify an analytic approximation for every supplied graph.

Existing graph-based code

E <- gseg_kmst(x = x, k = 3)
fit <- gseg1_graph(nrow(x), E)
fit <- gseg1(nrow(x), E)       # earlier API still works
fit <- gseg1(nrow(x), E = E)   # mixed named/positional calls also work

An explicitly supplied graph is used as supplied. k cannot be specified in that call because no graph is being constructed. gseg1_data() and gseg2_data() remain supported aliases; new code can call gseg1() and gseg2() directly. Repeated-observation scans remain available through gseg1_discrete() and gseg2_discrete().

The requested MST union is verified to have k * (N - 1) distinct edges. The counting bound k <= floor(N/2) alone does not guarantee greedy tree packing: a previous tree can exhaust a vertex. Such inputs now raise a clear construction error; choose a smaller explicit k or supply a graph. No silent fallback changes k or the similarity representation.

Pass a dist object through dissimilarity = ...; using it as x is rejected so that its geometry cannot silently change. Analytic integrations must report successful convergence and finite probabilities in [0, 1]. A failed analytic approximation raises an informative error; set pval.appr = FALSE and use justified permutation calibration rather than treating a failed approximation as a probability.

The package uses registered native kernels for exact MST construction, finite-data Euclidean distances and existing skew-correction graph counts. These implementations retain the same graph selection, tie handling and statistical formulas; floating-point contraction is disabled during compilation.

Reference compatibility

Raw observations use stats::dist() on every platform, as in the original CRAN package. Native graph construction and graph-moment optimizations remain active. Numerically reliable original root and tail calculations are retained; zero-skew/cancellation failures use stable equivalent arithmetic. Verified interval-index, permutation-diagnostic and validation repairs remain explicit behavior changes, so the complete release is not claimed output-identical to older versions on inputs affected by those defects.

Distance compatibility (1.2-4 validation candidate): the accelerated finite Euclidean path is restricted to exact verified R/stats build identities and passes an RNG-free arithmetic self-check. Other builds use stats::dist(). A new R installation may therefore retain reference speed until separately verified; no architecture-only arithmetic assumption is made. No parallelism or statistical parameters are changed.