Changelog

On this page

Changelog#

v0.1.0 (2026-02-15)#

Initial release.

Generators:

  • TVAR coefficient schedules: sinusoid, fourier, quasiperiodic, polynomial drift, logistic transition, multi-sigmoid, Gaussian bumps, smooth random

  • Time-Varying VAR(2) with rotating coupling matrices (TVVAR)

  • Lorenz system (RK4 integration)

  • Ornstein-Uhlenbeck process (exact analytical & Euler-Maruyama)

  • Colored noise family: white, pink (1/f), brown (1/f²), arbitrary spectral exponent

Embedders:

  • Delay embedding for state-space reconstruction

Models:

  • AR: classical OLS autoregressive model with AIC/BIC selection

  • TAR: two-regime threshold autoregressive model with bootstrap simulation

  • DeepLagEmbed: geometry-aware network for joint AR-order and time-varying coefficient estimation

  • ARMLP: MLP baseline for time-varying coefficient estimation

  • ARTransformer: Transformer encoder baseline for time-varying coefficient estimation

  • AnalyticalAR: non-learned least-squares AR baseline (PyTorch)

Utilities:

  • Composite loss functions: classification (CE), AR reconstruction (MSE), energy constraint, smoothness penalty, order regulariser

  • Training loop with ReduceLROnPlateau scheduling and gradient clipping

  • Benchmark loop with per-order accuracy, coefficient MSE, and signal MSE

  • Bicoherence for quadratic phase-coupling detection

Plotting:

  • Training history dashboard (2x6 grid with confusion matrix)

  • Per-order coefficient visualisation (true vs. predicted)

  • TVAR sample inspector (coefficients, signal, sliding-window power)

On this page