L-UQ 1.2.0

[1.2.0] — 2026-07-11

Changed (breaking)

• Python import namespace renamed lmoments → lmoments_uq to
  avoid collision with the unrelated lmoments and lmoments3
  packages already on PyPI, which also install a top-level lmoments
  module. The PyPI distribution name is unchanged (lmoments-uq); only
  the import changes: from lmoments_uq import .... Update any code
  that did from lmoments import .... Repository (L-UQ), distribution
  (lmoments-uq), import (lmoments_uq), and article all now align.
• Added Palaniappan Ramu as a second author in the package metadata
  (pyproject.toml, CITATION.cff), matching the article.

(Contains all 1.1.0 changes below; 1.1.0 was released on GitHub but
not published to PyPI, so 1.2.0 is the first PyPI release carrying the
bootstrap identification and population-truth benchmark.)

L-UQ 1.1.0

[1.1.0] — 2026-07-11

Adds uncertainty-aware identification and a population-truth benchmark,
in response to peer-review-style feedback on the manuscript.

Added

• identify_dist_bootstrap (Python) / Identify_dist_bootstrap.m
  (MATLAB): bootstrap-based, uncertainty-aware distribution
  identification. Resamples the data with replacement, re-identifies on
  each resample, and returns per-family selection frequencies, 95%
  percentile confidence intervals for (t3, t4), and a clear/ambiguous
  status flag. Addresses the fact that at small n the single "closest"
  family is often not statistically distinguishable from the runner-up.
• Bootstrap identification unit tests in all three suites (Python +5,
  MATLAB +3, Octave section G).

Changed

• The replication benchmark (replication/run_all.py) now scores every
  fit against the KNOWN parent distribution (population truth) rather
  than the small sample's own histogram: integrated absolute CDF error,
  extreme-quantile error, and the risk-relevant tail-probability error
  P(X > x_c) at the true 99th percentile, with Jensen-Shannon retained
  only as a secondary diagnostic. This removes histogram-binning
  sensitivity and the circularity of scoring a fit against the noisy
  sample it was estimated from.
• The benchmark additionally records identification accuracy (true
  family ranked first / in top three, split by 2-parameter point vs
  3-parameter curve families, and fallback rate), written to
  replication/output/identification_accuracy.csv.

L-UQ 1.0.1

[1.0.1] — 2026-07-09

Documentation-only release; no code changes.

Changed

• python/README.md (the PyPI project description): replaced the
  outdated "not validated by diffing against MATLAB output" caveat —
  written before a MATLAB installation was available — with the
  current verification status: MATLAB/Python equivalence to 1e-8 on
  fixed reference samples (via tests/octave_verify.m under GNU
  Octave 11.3 and tests/test_uq_matlab.m under MATLAB R2026a,
  19/19), plus machine-precision agreement with R's lmom where the
  closed forms coincide. Test count corrected (30), toolbox name
  updated to L-UQ.

L-UQ 1.0.0

[1.0.0] — 2026-07-08

First stable release, prepared alongside the Journal of Statistical
Software submission.

Added

• Python port of the full MATLAB toolbox (python/lmoments/):
  lmom, pwm, l_moment_ratios, identify_dist,
  parameter_estimation, parameter_identify, fit_best,
  pdf_l, cdf_l, random_l, kl_div, js_div.
• fit_best guarded fit with ranked fallback, in both languages
  (python/lmoments/parameters.py, fit_best.m): walks the
  ratio-diagram ranking and returns the first family whose closed-form
  estimator domain is satisfied, recording skipped families.
• Domain guards in both languages: estimator domain violations raise
  informative errors (ParameterEstimationError in Python,
  LUQ:... identifiers in MATLAB) instead of returning NaN.
• Explicit three-parameter Weibull support end-to-end in MATLAB
  (Parameter_estimation, PDF_l, CDF_l, Random_l); excluded
  from automatic identification by design (ratio-diagram curve overlap).
• Interactive Streamlit application (python/app.py).
• Test suites: 30 Python unit tests (python/tests/), a mirrored
  MATLAB suite (tests/test_uq_matlab.m), and an Octave-runnable
  verification script (tests/octave_verify.m, 38 checks) including
  machine-precision equivalence between the MATLAB and Python
  implementations on fixed reference samples.
• GitHub Actions CI (pytest on ubuntu/windows × Python 3.9/3.12;
  MATLAB suite via matlab-actions).

Fixed

• CDF_l.m: gamma branch now applies the same location shift as
  PDF_l.m (shifted-gamma PDF/CDF consistency).
• Identify_dist.m: round(x, 4) rewritten in portable form
  (round(x*1e4)/1e4) so the toolbox runs unmodified under GNU Octave.
• Identification tests: normal/gamma acknowledged as a degenerate pair
  (the zero-skew limit of the shifted gamma is the normal), same policy
  as Gumbel/GEV and uniform/GP.

Changed

• Toolbox renamed UQ → L-UQ (repository, paper, error identifiers)
  to be descriptive and avoid collision with the existing UQLab
  framework.
• MATLAB error identifier prefix UQ: → LUQ:.

Removed

• lhsgeneral.m (third-party utility with unconfirmed licensing);
  the repository is now 100% MIT. The README points to the original
  File Exchange entry.