Module pelt

PELT (Pruned Exact Linear Time) offline change point detection.

Detects structural breaks in embedding trajectories by finding timestamps where the mean vector shifts significantly.

Algorithm

1. Define cost function as sum of squared deviations from segment mean.
2. Dynamic programming with pruning: maintain set of candidate change points.
3. Penalty controls sensitivity (BIC default: dim * ln(n) / 2).

Complexity: O(N) amortized per Killick et al. (2012).

Structs

CumulativeSums 🔒

Cumulative sums for O(1) segment cost computation.

PeltConfig

PELT configuration.

Functions

build_cumsum 🔒

compute_segment_means 🔒

Compute mean vectors before and after a change point.

detect

Run PELT change point detection on a trajectory.

segment_cost 🔒

Cost of segment [start, end): sum of squared deviations from segment mean.

segment_mean 🔒