MSig 0.2.0

Changelog

All notable changes to MSig are documented here. The format follows
Keep a Changelog and the project adheres
to Semantic Versioning.

[0.2.0] — 2026-05-XX

Fixed

• 2D Gaussian rectangle CDF — vars_indep_time_markov in the
  gaussian_theoretical branch now uses the inclusion-exclusion formula
  F(b1,b2) − F(a1,b2) − F(b1,a2) + F(a1,a2). Affects every gaussian_theoretical
  result. Empirical and KDE branches are unchanged. None of the published
  case studies use gaussian_theoretical, so paper tables are not affected.
• Empirical conditional consistent normalization — numerator and
  denominator both use the lag-1 marginal over n−1 transition pairs.
  Removes a n/(n−1) bias that previously triggered the cond_p > 1 clamp
  on small series.
• KDE conditional uses lag-1 marginal — denominator now integrates
  gaussian_kde(y[:-1]) for theoretical consistency with the bivariate
  KDE numerator.
• set_significance always sets self.pvalue — every return path
  (including p_Q ∈ {0, 1} and n_matches ≥ max_possible_matches) now
  populates the attribute, eliminating discrepancy between return value
  and stored attribute.
• Three momenti scripts — NameError on undefined m (now s) and
  wrong relative data paths (../data/ → ../../data/). The MOMENTI
  experiments now run.
• scripts/compare_results.py schema — reads the actual CSV columns
  produced by experiment scripts (median_pvalue, median_probability,
  #sig_motifs(≤0.01), #sig_hochberg, significant as percentage)
  rather than the never-implemented pvalue/significant/pattern_probability columns.
• Population-density MOMENTI — adds s = 48 to motif lengths,
  matching paper Table 4 and the LAMA script.
• Logger consistency — logging.info → logger.info in MSig.py.
• __init__.py version drift — __version__ reads from
  importlib.metadata.version("msig"). No more manual sync with
  pyproject.toml.
• vars_dep_time_markov docstring — cites paper Eq. (4) for the
  unimplemented variable-dependent null-model formula.

Added

• pattern_prob_floor kwarg on Motif.set_significance — opt-in
  Laplace floor for p_Q = 0 cases (zero-frequency problem). Default
  None preserves 0.1.x behaviour.
• EXCLUSION_ZONE_FACTOR constant in every experiment script and
  experiments/common_utils.py — surfaces the trivial-match factor
  (default 0.5, matching published tables; paper §3.2 default is 0.25).
• AVERAGE_DELTA module constant — replaces the per-script
  average_delta = 0.3 hardcode.
• experiments.common_utils.get_dataset_paths — single-source helper
  for resolving data and results paths regardless of cwd; adopted by all
  nine experiment scripts.
• tests/golden/ — paper-reproducibility regression suite locks in
  the CSV outputs so future drift is caught.
• CITATION.cff, CONTRIBUTING.md, REPRODUCING_EXPERIMENTS.md — accessibility documentation.
• msig/py.typed — PEP 561 marker enabling downstream type checking.
• Validation that δ = 0 is rejected for kde and gaussian_theoretical
  null models (silently produced p_Q = 0 before).
• Private helpers _rect_prob_1d and _rect_prob_2d in msig/MSig.py
  centralise the rectangle-probability computation across model branches.

Changed

• bonferroni_correction(n_tests, alpha=0.05) — first parameter renamed.
  Backward-compatible: still accepts iterables for 0.1.x callers.
• idd_correction documentation — corrected to "identically-distributed
  dimensions" (was wrongly described as "Independent Dimension Discovery"
  applying BH FDR).
• mSTUMP "conservative" regime comment clarifies that the formula is
  D_max = √s · δ · 0.5 (which produced the published tables), not the
  paper's D_max = (1/q) Σ δ_j (no √s scaling).
• Python support range aligned to >=3.11,<3.14 across pyproject.toml,
  environment.yml, requirements.txt, classifiers, mypy, and black.
• bonferroni_correction and benjamini_hochberg_fdr test coverage —
  new canonical Benjamini & Hochberg (1995) §3 example replaces a
  non-standard "returns α when nothing significant" assertion.

Removed

• PYPI_DESCRIPTION.md — duplicate of README.md; never reached PyPI
  (pyproject.toml uses readme = "README.md"). Removed to eliminate
  drift between the two.
• Dead OverflowError fallback in set_significance — scipy.stats.binom.sf
  underflows silently and never raises this exception.

Migration from 0.1.x

• bonferroni_correction([0.1, 0.2, 0.3]) continues to work; new code
  should pass bonferroni_correction(3) directly.
• If any user code relied on gaussian_theoretical p-values, expect
  numerically different results after the inclusion-exclusion fix.
• If any user code relied on inspecting motif.pvalue after the
  p_Q ∈ {0, 1} edge cases, the attribute is now always correct.

[0.1.3] — 2025-10-27

• Hotfix release.

[0.1.2] — 2025-10-26

• Maintenance release.

[0.1.1] — 2024-07-03

• Initial PyPI publication; corresponds to paper-submission code.

[0.1.0] — 2024-07-03

• First release.