ntsa.tools

The primitives — one module per job; every function is re-exported at the package top level, so from ntsa.tools import delay_embed always works. The higher-level analyses built on these live one level up: ntsa.classification and ntsa.bifurcation.

At a glance

Module	What it does
runners	Fresh model instances (respawn) and long integrations past the transient (run_long).
embedding	Delay embeddings and lag/dimension selection (AMI, FNN).
maps	Return maps, Poincaré sections, peak-based transient/period detection.
geometry	Attractor geometry from pairwise distances: recurrence matrices, classical MDS, correlation dimension.
statistics	Power spectral density, autocorrelation, summary moments.
lyapunov	Lyapunov exponents: Benettin QR spectrum, leading exponent, Kaplan–Yorke, data-only Rosenstein.

Data-driven vs model-based

Needs	Functions
Data only (raw arrays)	Everything in embedding, maps, geometry, and statistics, plus rosenstein_lyapunov — they take a scalar series x or a trajectory Y, so they work on measurements with no equations. This is what ntsa.data.DataSeries builds on, and ntsa.classification is data-driven too.
Model equations	respawn and run_long (runners) and the rest of lyapunov — they instantiate and integrate a model implementing the protocol. So does ntsa.bifurcation (plot_bifurcation excepted, which just plots sweep output).

classify_regime accepts optional lam1/exponents evidence, which can come from the model-based lyapunov estimators or the data-only Rosenstein one.