.. -*- mode: rst -*-
.. vim:ft=rst

.. _changelog:

#############################
Regreg Development Changelog
#############################

The full VCS changelog is available here:

  https://github.com/regreg/regreg/commits/main

***************
Regreg releases
***************

Most work on Regreg so far has been by Brad Klingenberg (BK) and Jonathan Taylor (JT).

References like "pr/298" refer to github pull request numbers.

0.2.0 (October 2026)
====================

The first release from PyPI with the current API (PyPI had 0.1.3).

* Packaging: built with meson-python, with wheels for Linux, macOS (arm64) and
  Windows on CPython 3.10 to 3.14.
  Requires Python 3.10 or later, NumPy 2 compatible.
* Cox regression uses ``coxdev`` (>= 0.3.0) for the objective, gradient and
  Hessian (pr/92).
* Case weights for the GLM losses (pr/87).
* Documentation: a new site at https://regreg.github.io/regreg/ with
  executed examples and an API reference (pr/105, pr/106, pr/114).
* Bug fixes from two reviews of the package (gh-86, gh-108), including:

  - solvers: ``gengrad``, ADMM with a non-symmetric ``D``,
    ``container.solve(quadratic=...)``, ``FISTA.fit`` options (pr/95);
  - ``tfocs`` and ``nesta`` now agree with linear programming to about 1e-8
    (pr/107);
  - path algorithms: strong rules with unpenalized variables, the mixed lasso
    elastic net, cross-validation folds, ``basil`` stalls (pr/96, pr/112), and
    a tight KKT tolerance (``kkt_tol``) for variables entering the model, so
    paths agree with direct solves (gh-99, pr/115);
  - atoms, affine maps and smooth losses: projections, feasibility checks,
    offsets, sparse inputs, stable probit and cloglog losses (pr/97, pr/98,
    pr/109, pr/110, pr/111);
  - scikit-learn estimators (pr/101) and representations and doctests under
    NumPy 2 (pr/102).

* Tests run with pytest and use ``np.random.default_rng`` (pr/88, pr/100).
