Changelog#

Unreleased#

0.5.0rc1 - 2026-09-25#

Release candidate for testing before 0.5.0.

Added#

  • Added 21 correctness-gated pySHT-native methods, bringing the candidate API to 72 canonical public functions. The additions comprise Chen–Qin and Li high-dimensional mean tests; Chen–Zhang–Zhong covariance tests; energy and MMD distribution tests; distance covariance, HSIC, dHSIC, and distance multivariance; multivariate Henze–Zirkler and energy normality tests; EHY rectangular and simplex uniformity; alpha-energy compositional equality; and four circular-data procedures.

  • Added public pysht.independence and pysht.circular modules without top-level function re-exports, plus native-method API, chooser, glossary, formula-ledger, reproducible-audit, and installed-wheel coverage.

  • Added shared distance, metric-canonicalization, Gram-matrix, exact-orbit, kernel-bandwidth, and multiplier-plan infrastructure. Built-in kernel choices are deliberately limited to RBF and Laplacian kernels with a fixed positive or observed-median bandwidth.

Changed#

  • Renamed mean_variance.as_1samp to lrt_1samp, mean_variance.zxc_2samp to exact_lrt_2samp, and both Lee–You–Lin lyl_2samp functions to maximum_pairwise_bayes_factor_2samp. These are direct pre-1.0 breaking changes; no callable aliases are provided.

  • Extended the API organization with native categories [11] Independence and [12] Circular data, while keeping the R migration crosswalk limited to SHT identities and recording new research methods in a separate catalog.

Corrected#

  • Corrected Johansen’s finite-sample F adjustment, including its exact reduction to the squared Welch test in one dimension.

  • Included floating-point permutation ties in the LJW and Thulin upper tails, and rejected rank-deficient Hotelling inputs using equilibrated observation matrices instead of relying on Cholesky success alone.

  • Treated saturated multivariate normality samples (n = p + 1) as the degenerate all-ties case with p-value one and an explicit diagnostic.

  • Limited overflow fallback to affected features so normality and mean tests preserve representable small variations in other coordinates.

  • Preserved Schott (2001) results across independent changes of feature units with observation-space SVD whitening and rank checks.

  • Preserved representable variance-test tail probabilities when their F or chi-square pivots overflow or underflow in float64.

  • Bounded metric canonicalization work and accepted duplicate compositions for Monte Carlo alpha-energy inference beyond small exact orbits. Bitwise seeded geometric replay is conditional on completing canonicalization; uniform randomization remains statistically valid after its fallback.

  • Identified the Chen–Qin variance implementation as the Li–Chen unbiased trace-estimator variant and distinguished simulation release gates from exact finite-sample calibration, with binomial intervals in native audits.

  • Made distance-, kernel-, independence-, circular-, and compositional statistics invariant to the transformations implied by each method and stabilized seeded resampling across tractable symmetric geometries and ties.

  • Centered Biswas–Ghosh coordinates before numerical scaling so exact permutation inference preserves representable within-sample geometry near a huge common location instead of collapsing ULP-scale separations.

  • Stabilized the alpha transformation continuously at alpha=0, including extremely small positive and negative values, while preserving the structural-zero rule for positive alpha.

  • Made simplex domain validation and alpha-energy canonicalization invariant to a common component permutation at floating-point tolerance boundaries.

  • Aligned circular Monte Carlo tie handling so an identical seeded null draw is always included in its own Rayleigh, Watson, or Hermans–Rasson upper tail despite harmless preprocessing-order roundoff.

  • Removed the unsupported unique-minimum restriction from Li’s multi-sample test, added literal tied-reference behavior, and validated a balanced three-group null design through a replayable 20,000-dataset scenario.

  • Stabilized Srivastava–Du and the Yao, Krishnamoorthy–Yu, Johansen, and Nel–Van der Merwe mean tests for coordinates with radically different units, while replacing avoidable p-by-p trace, projection, and subspace work with observation-space equivalents where applicable.

  • Preserved representable extreme Pearson–Neyman lower-tail probabilities when the likelihood ratio itself underflows, and bounded CLX entrywise variance evaluation so it no longer materializes an n-by-p-by-p tensor.

  • Evaluated Student-t, chi-square, F, and Clopper–Pearson upper quantiles through survival functions so confidence bounds remain finite and representable when the requested confidence level is adjacent to one.

  • Made shared resampling calibration reject NaN statistics explicitly and apply identical scalar and batched upper-tail ordering at positive and negative infinity.

  • Stabilized the joint mean/covariance likelihood ratio for full-rank columns spanning hundreds of orders of magnitude, and removed unnecessary p-by-p identity Cholesky allocations from the default CZZ and LLZS paths.

  • Replaced global-scale covariance definiteness checks with diagonal equilibration and entrywise symmetry checks, preserving valid null matrices whose marginal variances span the float64 exponent range.

  • Stabilized Chen–Zhang–Zhong identity statistics when every coordinate is scaled to an extreme finite magnitude, reconstructing second- and fourth-order terms analytically instead of overflowing or underflowing intermediate powers.

  • Made CI and release validation measure branch coverage explicitly and force a fresh, complete warning-as-error documentation build.

Withheld by scientific gates#

  • Kept Xue–Yao xy_2samp private because the planned 999-draw calibration missed its audited null-size gate.

  • Kept Yu–Li–Xue ylx_2samp private until the joint component-independence and 20,000-dataset combination gate pass. JWJWZ covariance and YLXL joint procedures have no public callable because their equation and calibration audits remain incomplete.

  • Kept Ball Divergence private because the optimized floating-distance rank engine cannot yet distinguish every representable near-tie while preserving exact closed-ball symmetries; one adversarial case changed the literal exact p-value.

0.1.0 - 2026-08-10#

Added#

  • Established the Python 3.12+ package, NumPy/SciPy runtime, C++17 native kernel, stable-ABI wheel, and typed-package scaffold.

  • Implemented and exposed 51 canonical Python functions representing 52 of the 54 public statistical routine identities in SHT 0.1.9. The catalog now covers univariate and multivariate means, variances, covariances, joint mean and variance, joint mean and covariance, equality of distributions, normality, rectangular uniformity, and simplex uniformity.

  • Added immutable StatisticalTestResult, HypothesisTestResult, ResamplingTestResult, DistanceTestResult, and BayesFactorTestResult contracts. Frequentist results render in an R htest-style report; Bayesian results retain component log Bayes factors without fabricating a p-value.

  • Added exact-enumeration and corrected Monte Carlo infrastructure, including (b + 1) / (B + 1) p-values, exceedance counts, conditional Monte Carlo standard errors, and 95% binomial tail-probability intervals.

  • Added shared validation, local random-number handling, stable whitening and linear algebra, log-domain tail and Bayes-factor calculations, bounded optimization, and overflow-aware scaling utilities.

  • Added the complete 56-export R migration crosswalk, searchable method-name glossary, category-organized API reference, interpretation-centered user guide, and method-specific validation ledgers.

Changed#

  • Standardized every public function on lowercase snake_case author tokens and scientific category modules. Test functions are not re-exported from top-level pysht.

  • Renamed the equality-of-distributions entry point directly to pysht.equaldist.bg_2samp; no obsolete biswas_ghosh_2samp alias is provided.

  • Mapped both mvar1.1998AS and mvar1.LRT to the single algebraically identical pysht.mean_variance.as_1samp implementation.

  • Withheld mean2.2014CLX from the public API after its practical-size Gumbel calibration failed the release gate. Its implementation and validation ledger remain private so the unresolved scientific limitation is visible.

  • Withheld cov1.2012Fisher because its earlier optimized null-calibration evidence was not reproducible through the public path and practical fresh public-path regimes did not pass. The corrected implementation remains a private validation target.

  • Made the Lee–You–Lin procedures return maximum and component log Bayes factors. Their published default shrinkage is derived from alpha=2.01, sample size, and dimension, with an explicit gamma override.

  • Replaced the invalid Biswas–Ghosh asymptotic branch with exact or corrected Monte Carlo permutation calibration.

  • Switched release-wheel construction to cibuildwheel so Linux artifacts are repaired to portable manylinux wheels while preserving the CPython 3.12+ stable ABI.

Corrected#

  • Repaired known SHT defects in null-covariance whitening, Li–Chen scaling, two-sided Wu–Li covariance tests, Pearson–Neyman and Muirhead rejection tails, Cao–Park–He group indexing and variance branches, adjusted and robust Jarque–Bera defaults, and two-sample Dempster calibration.

  • Held random projection and subspace plans fixed across observed and permuted statistics, canonicalized seeded plans across row and group order, and avoided touching NumPy’s global random state.

  • Stabilized Zhang–Xu–Chen quadrature and roots, extreme-value tails, likelihood ratios, covariance traces, and Bayes-factor kernels for extreme float64 magnitudes.

  • Centered samples against null or shared anchors before scaling throughout the mean, variance, covariance, normality, and joint-parameter paths, with explicit opposite-endpoint fallbacks. This preserves representable variation near very large common locations and across heterogeneous scales.

  • Corrected the Yang–Modarres Q3 asymptotic law for the nonzero correlation of its signed components, retained the published robust Jarque–Bera denominator constant 64 instead of SHT’s erroneous 24, and made the Arnold–Shavelle calculation fully log-domain.

  • Tightened exact-resampling count consistency, Bayes-factor component/max consistency, simplex-interior validation, optimizer convergence checks, and covariance eigensolver and extreme-tail boundaries.

Validation and release engineering#

  • Added primary-paper formula ledgers, literal or trusted numerical oracles, fixed fixtures, transformation invariants, boundary tests, null-calibration audits, named-seed targeted alternative-power checks for every public frequentist procedure, Bayesian evidence-direction checks, and explicit legacy-correction records.

  • Added strict typing, linting and formatting, branch-coverage enforcement, warning-as-error Sphinx builds, source-distribution checks, cross-platform portable wheel builds, and isolated installed-wheel smoke tests for the exact 51-function public surface on Python 3.12–3.14.

  • Added a release-triggered, cross-platform GitHub Actions workflow for tokenless PyPI Trusted Publishing. The workflow assembles one source archive and four platform wheels, smoke-tests the exact artifacts, records SHA-256 hashes, and grants OIDC permission only to the final protected publishing job.

  • Added a minimum-supported NumPy/SciPy CI gate, replayable release-audit runners, and the BSD/MIT notices required by the native extension’s statically linked nanobind and tsl::robin_map code.