Changelog¶

Unreleased¶

Self-checking teaching notebooks (documentation only, no library changes): two executed notebooks now walk through the mathematics and the workflow of each backend, with every derivation followed by an assertion and a closing pass/fail audit: examples/wood_backend_math_tutorial.ipynb and examples/eilers_backend_math_tutorial.ipynb. They are linked from Quick Start, Examples, Mathematical Foundations and the backend math chapters.

Released 2026-10-18

Reference Parity with R ``mgcv`` & ``pyGAM``: Full numerical and mathematical alignment against reference implementations.
Exact Integrated Derivative B-Spline Penalties: Upgraded Wood backend bs basis penalty to Gauss-Legendre quadrature of integrated squared derivatives matching R’s smooth.construct.bs.smooth.spec.
REML Null Space Dimension: Corrected REML calculation of unpenalised null space dimension \(M_p = p - \operatorname{rank}(S_{\text{total}})\) and added coarse-grid seeding before L-BFGS-B optimization.
Active-Set Shape Constraints: Enforced pyGAM active-set asymmetric penalty formulation with \(\lambda_{\text{con}} = 10^9\), \(\lambda_{L2} = 10^{-3} I\), and exact periodic basis column wrapping.
Verification & Test Suite: 55 automated tests verifying numerical tolerances against R mgcv and mathematical invariants.

Released 2026-10-17

Updated documentation styling and replaced em dashes across all guides and examples.
Added comparative overview with R mgcv and pygam.
Enhanced package keywords, canonical metadata, and repository discoverability.

Released 2026-10-15

Wood (2011) backend: Native implementation of Simon Wood’s GAM methods supporting thin plate splines ('tp', 'ts'), cubic regression splines ('cr', 'cs'), cyclic cubic splines ('cc'), P-splines ('ps', 'cp'), B-splines ('bs'), random effects / ridge ('re'), and tensor products ('te').
Automatic REML smoothing optimization: Smoothing parameters are solved continuously via Newton-Raphson / BFGS minimization of REML or GCV criteria without needing discrete grid searches.
Summary tables: model.summary() provides parametric coefficients, effective degrees of freedom (EDF), approximate F-tests, and p-values matching Wood (2013).
Symmetric dual-engine architecture: Symmetrically organized into magli.eilers and magli.wood with unified LinGAM and GAMCore factory dispatchers.

Shared ``_GAMBase`` lifecycle: magli._base now holds the code both engines inherit (predict, partial_dependence, confidence/prediction intervals, loglikelihood/aic/bic/ deviance_explained, and all decomposition/diagnostic plotting). Note the scale convention it abstracts, where Eilers-Marx stores scale as σ (uses s**2 as variance) and Wood stores it as σ² directly.
Observation weights across both backends: fit(x, y, weights=w) and gridsearch(x, y, weights=w) scale observations by \(\sqrt{w_i}\) in both the Eilers-Marx and Wood backends (replacing the earlier robust=True IRLS option).
Cyclic P-splines (``’cp’``): dedicated basis with wrapped knots and a wrapped difference penalty.
Thin-plate shrinkage (``’ts’``): now charges the null-space (smallest) eigenvalues as in Wood (2017), instead of the largest ones.
Fixed R² computation: Corrected deviance explained calculation against the sample mean null deviance.
Input validation: Enforced NaN and Inf validation across all fit methods.
Extrapolation: Explicit linear extrapolation for predictions beyond boundary knots.

Released 2026-05-01

Shape constraints for s() terms and LinGAM: mono_inc, mono_dec, convex, concave, periodic.
Fast path is now the only path: removed fast parameter from gridsearch(). All grid searches use QR+Cholesky by default.
Robust grid search uses dedicated IRLS path with threaded evaluation.
Statistics methods added to both GAMCore and LinGAM: loglikelihood(), aic(), bic(), deviance_explained().
Selective grid search (via search_terms parameter): restrict tuning to specific term indices or types.

Modular src/magli/ package layout with __slots__ for memory efficiency.
PEP8 naming: b_spline_basis, solve_pirls, irls_solve, etc.
b_spline_basis supports periodic wrapping.
Penalty builders split into dedicated module with new first_diff_penalty and circular_second_diff_penalty.

Released earlier

Initial release with LinGAM (single-term) and GAMCore (multi-term formula interface).
B-spline basis via vectorised De Boor recursion.
SVD data-augmentation solver (PIRLS).
Huber IRLS robust fitting.
GCV grid search with QR+Cholesky fast path and SVD basic path.
Partial dependence, confidence/prediction intervals, and decomposition plotting.