ESST

ESST entangles past and future Hankel matrices before decomposition and scores how their contributions differ. It is a one-dimensional research method intended for changes in local signal structure.

Use the tuning guide for a parameter-by-parameter workflow and runtime advice.

Parameters

window_length : int

Subsequence length in samples and main temporal scale.

n_windows : int, optional

Number of subsequences per side. Defaults to window_length // 2.

lag : int, optional

Separation between compared matrices. Defaults to n_windows.

rank : int, default 5

Number of singular directions used by the score.

scale : bool, default True

Min-max scale the signal to [1, 2].

method : {"fbrsvd", "rsvd"}, default "fbrsvd"

Randomized decomposition implementation. Use "rsvd" with fast Hankel products.

random_rank : int, optional

Sampled randomized-SVD dimension. Defaults to at most rank + 10.

scoring_step : int, default 1

Distance in samples between evaluated positions.

use_fast_hankel : bool, default False

Use implicit FFT-based Hankel products. Requires method="rsvd".

mitigate_offset : bool, default False

Remove column-wise offsets. Cannot be combined with fast Hankel products.

Methods

transform(time_series)

time_series : numpy.ndarray, shape (n_samples,)

Signal to score.

Returns

score : numpy.ndarray, shape (n_samples,)

Entangled subspace-change score with an initial zero-padded region.

Minimal Example

import numpy as np
from changepoynt.algorithms.esst import ESST

signal = np.r_[np.sin(np.linspace(0, 8 * np.pi, 200)),
               np.sin(np.linspace(0, 20 * np.pi, 200))]
score = ESST(window_length=40).transform(signal)

Reference: Boelter et al. (2025).