Skip to content

Prepare v2.3.0 release - #148

Merged
HadenSmith merged 79 commits into
mainfrom
release/2.3.0
Nov 10, 2026
Merged

HadenSmith merged 79 commits into
mainfrom
release/2.3.0

Conversation

@HadenSmith

@HadenSmith HadenSmith commented Nov 10, 2026 •

Copy link
Copy Markdown
Collaborator

Summary: metadata updates

  • Numerics/Numerics.csproj: Version 2.3.0, AssemblyVersion 2.3.0.0, and PackageReleaseNotes for 2.3.0.
  • CITATION.cff: version 2.3.0 and date-released 2026-11-10.
  • codemeta.json: version 2.3.0, dateModified 2026-11-10, and runtimePlatform .NET 10.0 and .NET Framework 4.8.1.
  • .github/workflows/Snapshot.yml: snapshot version advanced to 2.3.1.
  • python/pynumerics/src/pynumerics/_version.py: __version__ 2.3.0, NUMERICS_VERSION 2.3.0 and NUMERICS_ASSEMBLY_VERSION 2.3.0.0. The SHA-256, source commit and build command stay empty; the wheel build writes the published package's identity.
  • python/pynumerics/CHANGELOG.md: the 2.3.0 entry, also rendered in the Sphinx docs.
  • README.md, docs/index.md and docs/getting-started.md: supported frameworks (net10.0 and net481), the 2.3.0 install command, the 2.3.0 feature overview, and the pynumerics install.
  • python/pynumerics/pyproject.toml: description, keywords, Python 3.10 to 3.14 and Linux and Windows classifiers, and documentation and changelog URLs.
  • Package validation stays on PackageValidationBaselineVersion 2.2.0 for this release. Its only suppressions are the two PKV006 entries for the dropped net8.0 and net9.0 targets. After 2.3.0 is published, the baseline advances to 2.3.0 and those suppressions are removed.

Summary: user-facing changes

  • Breaking:
    • Targets net10.0 and net481 only. .NET 8 and .NET 9 reach end of support on November 10, 2026; projects on those runtimes can stay on 2.2.0.
    • GeneralizedExtremeValue.ParametersFromLinearMoments throws for λ2 ≤ 0, non-finite λ1 or λ2, or |τ3| ≥ 1.
    • Dijkstra and Network reject negative edge weights.
    • Deterministic.PDF, CDF and InverseCDF throw for a NaN or infinite Value.
    • Tools.LogSumExp returns +inf for +inf input and -inf for an empty list.
    • GoodnessOfFit.PBIAS follows the Moriasi et al. (2007) sign.
    • pynumerics replaces the unpublished rmc-numerics package, with no aliases.
  • Fixes:
    • StudentT tails at large ν, Beta.Incomplete and its inverse, and the inverse incomplete gamma.
    • Corrected two-pass variance, covariance and Pearson correlation.
    • Joe, Clayton, Frank and Gumbel copula densities, InverseConditionalCDF edges, and AMH at θ = 0 and θ = 1.
    • GEV, Gamma and Pearson Type III L-moments, and GPA method of moments.
    • Global KDE and Truncated modes, the EmpiricalDistribution NormalZ NaN, and the TruncatedDistribution XML round trip.
    • GoodnessOfFit.KolmogorovSmirnov and AndersonDarling compute on a sorted copy, with new dataIsSorted overloads.
    • Linear Fourier autocorrelation, regression prediction intervals and the RungeKutta Fehlberg error term.
    • AdaptiveSimpsonsRule reports non-finite results as failures; Powell, MultiStart and MLSL stop cleanly at the evaluation budget.
    • USGS Water Data API v1, non-numeric USGS peaks skipped, and GHCN snowfall units.
  • Additions:
    • Regional frequency analysis (Numerics.Distributions.Regional).
    • GaussianProcessRegression, Kriging, ElasticNetRegression and PrincipalComponentAnalysis.
    • LBFGSB, CMAES, DynamicallyDimensionedSearch, ParetoFront, DynamicProgramming and Knapsack, with optimizer progress and cancellation.
    • The QL eigensolver, opt-in QAGS, relative-tolerance Brent.Solve and Dijkstra.ShortestDistances.
    • Column statistics and plotting-position percentiles.
    • Compensated sums, LogSumExp by axis, and Poisson and Binomial array members.
    • The pynumerics Python package, published to PyPI automatically after the NuGet release.
  • Platforms and CI:
    • The net10.0 build and pynumerics are tested on Linux in the .github/docker/linux-test image (Ubuntu 24.04, .NET 10 SDK, Python 3.12).
    • Bitwise test references stay exact on Windows; on other platforms each test uses the smallest tolerance measured on Linux, because .NET math functions call the platform C library.
    • PythonWheel.yml builds the wheel from the PR source and tests it on Linux, in the Docker image and on Windows. PythonPublish.yml runs after NuGetPublish.yml succeeds for a release.
  • Performance:
    • The QL eigensolver.
    • Poisson arrays (faster than scipy.stats.poisson.ppf from n = 200).
    • Dijkstra (faster than SciPy csgraph at 1e4 to 1e6 nodes in .NET).
    • Tools.Expm1, column statistics and pynumerics marshalling.
  • Results that move are listed in the release notes. Bulletin 17C conditional moments and the default AdaptiveGaussKronrod path are bitwise unchanged.
  • Tests: unit tests with mpmath, SciPy, NumPy, R and published reference values cover each addition and fix: 4,075 .NET tests on net10.0, 4,058 on net481, and 5,770 pynumerics tests.

Validation

  • dotnet restore
  • dotnet build -c Release: 0 warnings, 0 errors (also with --no-incremental)
  • dotnet test -c Release --no-build -f net10.0 --filter "TestCategory!=LiveServiceIntegration" with VSTEST_CONNECTION_TIMEOUT=600: 4,075 passed, 0 failed
  • dotnet test -c Release --no-build -f net481 --filter "TestCategory!=LiveServiceIntegration" with VSTEST_CONNECTION_TIMEOUT=600: 4,058 passed, 0 failed
  • dotnet pack Numerics/Numerics.csproj -c Release -p:Version=2.3.0 --no-build -o ./packages: package validation against 2.2.0 clean apart from the PKV006 suppressions for net8.0 and net9.0
  • Inspect the .nupkg metadata: version 2.3.0, lib/net10.0 and lib/net481 only, and the 2.3.0 release notes
  • pynumerics: ruff check clean, pytest against this build: 5,770 passed, 0 failed (5 skipped, 3 benchmarks deselected), sphinx-build -W clean with the changelog page
  • Linux Docker check (.github/docker/linux-test): Test_Numerics net10.0 passed with no failures, ruff clean, pytest against the wheel 5,769 passed, 0 failed (6 skipped, 3 benchmarks deselected), and the fresh-container identity check passed (CoreCLR, assembly 2.3.0.0, SHA-256 matches)
  • The three live USGS examples in docs/data/time-series.md run against the USGS Water Data API
  • The PythonWheel.yml check on this PR builds Numerics.dll from the PR source, then builds and tests the wheel on Linux (ubuntu-24.04 and the linux-test Docker image) and Windows
  • After the GitHub Release is published, NuGetPublish.yml succeeds and PythonPublish.yml starts by itself: it waits for nuget.org, builds the wheel from the published package with its SHA-256 written in, tests it on Linux and Windows, and publishes to TestPyPI
  • Approve the pypi environment to publish to PyPI
  • pip install pynumerics==2.3.0 in a clean environment with the .NET 10 runtime; pynumerics.numerics_identity() reports assembly version 2.3.0.0 and the SHA-256 of lib/net10.0/Numerics.dll in the published package, and pynumerics.assert_numerics_identity() passes

Append the graphify section (query, path, and explain before raw search; graphify update after code changes). The rules apply only when a local graphify-out/graph.json exists.
…one and XML

Clone rebuilt the table through the two-argument constructor, so a survival
(descending) or tied-probability table cloned into an invalid ascending,
strictly increasing one. Clone now copies the orders, strictness and smart-search
setting. ToXElement writes StrictXValues and StrictProbabilities only when they
depart from the default, so existing documents stay byte-identical, and
FromXElement reads them. Adds the StrictXValues and StrictProbabilities
read-only properties.
…nd typed wrappers

python/rmc-numerics packages lib/net10.0/Numerics.dll from the RMC.Numerics 2.2.0
NuGet package byte for byte (SHA-256 checked by scripts/build_wheel.py) into one
py3-none-any wheel whose version equals the NuGet version. rmc_numerics.runtime is
the single CoreCLR loader (refuses a prior bare import clr, resolves DOTNET_ROOT,
memoized per process, add_assembly for dependent assemblies); marshal holds the
bit-exact bulk copies; identity reports the loaded assembly's version and SHA-256.
The wrappers call the Numerics array overloads and generators for distributions,
copulas, paired data, statistics, sampling designs, Brent, Tools.Expm1/Log1p,
UncertaintyAnalysisResults and BootstrapAnalysis.
Bound scalar overloads (pythonnet resolves an overloaded method group on every call, which
dominates the cheap scalar members); LogPDF/LogCDF/LogCCDF; type name, parameter names,
SetParameters, ValidateParameters messages and the IStandardError covariance; per-parameter-set
quantiles, validity and member arrays for ensembles held as rows; create_distribution through
UnivariateDistributionFactory, Uniform, Mixture weights and components; Linear.Extrapolate;
Statistics.LinearMoments; StratificationBin records and XValues over several ranges;
BrentSearch, Beta.Incomplete and IncompleteInverse, Matrix products, MatrixRegularization,
the Cholesky factor and LinkController; UncertaintyAnalysisResults accepting a .NET member
array.
…rameter construction

errors.net_errors re-raises the Numerics argument and arithmetic exception families as
ValueError, ArithmeticError, NotImplementedError and RuntimeError, so callers catch them
without importing System. BootstrapAnalysis.members and parameters keep a large replicate
ensemble on the .NET side; has_parameter_covariance reports IStandardError; LnNormal gains
from_log_parameters and direct_method_of_moments.
A deep copy of a distribution wraps its Clone and a deep copy of Linear rebuilds the interpolator
from its knots, so copy.deepcopy of a structure holding wrappers works (pythonnet objects cannot be
pickled) and the copies share no search cursor.
pythonnet resolves an overloaded method group on every call, and Marshal.Copy has more than a dozen
overloads: resolving them cost about 170 microseconds per transfer, twenty times the copy of a short
array. The exact overloads and the typed array constructors are now bound on first use; every
transfer copies the same bits.
…alidity over many parameter sets in one call
…e: quantile matrix and validity in one call)
…, assembly 2.2.0.0, source 279a5a4) and stage it with --dll
…ric construction, moments, density, GenerateRandomValues bitwise the .NET draws) with tests and the 2.3.0 notes
…_longs/to_numpy_longs, to_net_sbytes/to_numpy_sbytes, to_net_bools/to_numpy_bools)
…bind_list) instead of resolving the method group per call
…sum w F / sum w; equal weights reproduce the unweighted curve bit for bit)
…the finite entries, bitwise Statistics.Percentile per column (rmc-numerics statistics.column_percentiles)
…a known location (rmc-numerics GeneralizedPareto.fit_known_location)
…remeValue (probability transform once per column, shared InverseCDF kernel, bitwise identical to the per-row loop)
… reproduce Function(n, alpha) bit for bit (rmc-numerics statistics.weighted_plotting_positions)
…man-Fan type 6, Hazen type 5, Kaplan-Meier type 4; rmc-numerics statistics.plotting_position_percentile)
…a single unit weight, null members, a large ensemble's reproducibility and the range arguments
…llel reproducibility over 300 columns, k below zero
…ver two observations, observations at the location and the exponential fallback start
…g instance unchanged (also when invalid), closed-form InverseCDF checks
Drop the net8.0 and net9.0 targets from the library and test projects
(.NET 8 and .NET 9 reach end of support on November 10, 2026). Enable
pack-time package validation against RMC.Numerics 2.2.0, supply the
System.ValueTuple facade to the net481 comparison, and record the
suppressions for the two dropped frameworks. Link the new oracle and
time-series fixtures into the test project, and ignore internal working
notes under docs/reviews and docs/release-notes.
Add Tools.CompensatedSum and CompensatedCumulativeSum (Neumaier by
default, Kahan on request), the KahanAdd, NeumaierAdd and NeumaierResult
steps, LogSumExp over a matrix axis, Hypot, and list forms of Expm1 and
Log1p, and a range-partitioned parallel loop helper. LogSumExp follows
scipy.special.logsumexp, including the +inf and empty-list results.
Log1p uses the compensated form at every finite argument above -1, and
Expm1 takes the direct difference for |x| > 1/2. Tests cover bitwise
references, accuracy sweeps and the JSON converters.
…amma

Beta.Incomplete uses the BGRAT and BUP expansions of Algorithm 708 where
the continued fractions converge slowly, IncompleteInverse refines small
tails in log space, and LogFunction uses BETALN, ALGDIV and BCORR. One
bracketed Halley solver serves the inverse incomplete gamma for every
shape and probability, and Trigamma is accurate to full precision. Add
Bessel.I0e, Bessel.I1e and Evaluate.Chebyshev. Tests compare against an
mpmath oracle grid.
Add EigenSolverMethod with a QL path (SIMD tred2 and blocked tql2,
parallel from n = 128, bitwise equal to the sequential EISPACK
procedures) alongside Jacobi. CholeskyDecomposition uses a blocked
kernel for n >= 16 and gains the non-throwing TryCholesky with diagonal
jitter, and Matrix gains TransposeMultiply. Tests cover NumPy
references, ill-conditioned inputs and the unchanged small-matrix path.
Add Statistics.Median (used by MatrixRegularization in place of a
private median), column means, variances, standard deviations, L-moments
and percentiles, column goodness-of-fit against Uniform(0, 1), plotting-
position and weighted percentiles, general-order L-moments, the exact
two-sample Kolmogorov-Smirnov test, the Hartigan dip test and an
uncentered autocorrelation. Variance, covariance and Pearson correlation
use corrected two-pass sums. Regression prediction intervals include
leverage, and PBIAS follows Moriasi et al. (2007). TimeSeriesDownload
uses the USGS Water Data API v1, reads GHCN snowfall in millimeters and
gains parsers for saved responses. Tests use R and SciPy oracles and
saved service responses.
…utta fix

AdaptiveGaussKronrod adds QUADPACK QAGS through the Method property; the
default local bisection path is bitwise unchanged.
Fourier.Autocorrelation is now the linear autocorrelation, and Fourier
gains Convolve and BluesteinFFT. The RungeKutta Fehlberg error term
weights k5, and Cash-Karp steps use the fifth-order solution. Tests
cover QUADPACK reference integrals, convolution and ODE step control.
…ports

Add LBFGSB (L-BFGS-B 3.0), CMAES with IPOP restarts,
DynamicallyDimensionedSearch, ParetoFront, DynamicProgramming and
Knapsack, plus cancellation and progress reporting on every Optimizer.
Brent.Solve ports SciPy's brentq with a relative tolerance, and
BrentSearch ports SciPy's bounded minimizer. Dijkstra adds multi-source
CSR ShortestDistances, and Dijkstra and Network reject negative edge
weights. Tests compare against SciPy, pycma and generated oracles.
Add list overloads of the IUnivariateFunction and ILinkFunction members
as extension methods, bitwise equal to the scalar members.
CompositeFunction caches its per-child locks, and EnsembleFunction
parses its template once and pools the parsed elements; outputs are
unchanged. Link functions document their clamp bounds and exceptions.
Tests cover the extensions and exact equality with the prior outputs.
…tstrap

Add the opt-in NUTS LogLikelihoodAndGradientFunction, cancellation and
the current chain index on MCMCSampler, progress and cancellation on
Bootstrap, and LatinHypercube overloads that take a generator. NUTS
fixes its smoothed step size before sampling, SNIS takes the full log-
posterior and resamples by the inverse weight CDF, random initialization
uses a seeded generator, and an exact-zero uniform no longer accepts
every Metropolis proposal. Bootstrap and SNIS run sequentially and
bitwise equal to the parallel run when parallelism is off. Tests pin
seeded reference streams and cover the diagnostics and the bootstrap.
Correct the StudentT tails at large degrees of freedom, the GEV, Gamma
and Pearson Type III L-moments, and the generalized Pareto method of
moments. KernelDensity and TruncatedDistribution modes search for the
global peak. Add per-element Poisson and Binomial array members, the
Poisson-GEV and GPD conversions, GeneralizedPareto.MLE with a known
location, ParameterSetQuantiles, and list forms of CCDF and the log
functions. GEV ParametersFromLinearMoments throws for a non-positive
L-scale. Tests compare against SciPy and mpmath survival and quantile
oracles.
…t CDF

Evaluate the Joe and Gumbel densities in log space, fix Clayton for
negative theta and the Frank overflow, use closed-form Gumbel and Joe
conditional CDFs, and return exact InverseConditionalCDF limits. Add
list forms of the copula PDF, CDF and conditional functions,
Multinomial.CategoricalIndices, and a Genz-Bretz MultivariateStudentT
CDF with error controls. Tests cover reference densities, limits and
seeded draws.
Add the Numerics.Distributions.Regional namespace: RegionalData (site
L-moment ratios, regional averages and discordancy),
RegionalHomogeneityTest (H1, H2, H3 and Z measures from seeded, thread-
count-independent kappa simulations) and RegionalGrowthCurve (growth
factors and site quantiles). Tests reproduce published Hosking-Wallis
values.
…nd PCA

Add GaussianProcessRegression with squared exponential, Matern and
exponential ARD kernels, Kriging with variogram models,
ElasticNetRegression (glmnet covariance-update coordinate descent with
k-fold cross-validation) and PrincipalComponentAnalysis on the QL
tridiagonal path. Tests compare against scikit-learn, DiceKriging,
gstat, glmnet and prcomp oracles, kept with the R and Python scripts
that generate them.
Add Benchmarks_Numerics with benchmarks for the eigensolver, linear
algebra kernels, integration, optimizers, Dijkstra, distributions,
copulas, statistics, machine learning and LogSumExp, plus the sweep
runner used for bitwise reference checks. Add the project to the
solution.
Add python/pynumerics, which ships the Numerics.dll from the NuGet
release (SHA-256 checked) with one CoreCLR loader per process, bulk bit-
exact NumPy marshalling, Python exceptions that carry the Numerics
messages, and typed wrappers over every Numerics area. It replaces the
unpublished rmc-numerics package. Includes the Sphinx guides and the
pytest suite.
Document regional frequency analysis, Fourier methods, the bootstrap,
the new optimizers, the eigensolver and integrator options, machine
learning, column and plotting-position statistics, the new hypothesis
tests and the time-series download changes, and add the corresponding
references. The README states the net10.0 and net481 targets,
CONTRIBUTING adds the fast-unit-test guideline, and the contributor
guidance adds the pynumerics coverage rule and validation steps.
Install only the .NET 10 SDK in the integration, release and NuGet
workflows, run package validation against the baseline release in
integration, and let the release workflow retry an existing tag and skip
live-service tests. Build and test the pynumerics wheel, and add
PythonPublish for TestPyPI and PyPI. Update the issue template examples
to net10.0.
Set Version 2.3.0, AssemblyVersion 2.3.0.0 and the 2.3.0
PackageReleaseNotes, update CITATION.cff and codemeta.json to version
2.3.0 with the release date and supported runtimes, and advance the
snapshot version to 2.3.1.
…cs, wheel, pytest and identity checks; shell scripts keep LF endings
…ion and SHA-256 into the wheel's _version.py; _version.py declares the 2.3.0 release values
…ilds from the nuget.org package; the PR wheel ships a Numerics.dll built from the PR source
…e; on other platforms they use the smallest ulp tolerance measured on Linux, or Monte Carlo error where the NUTS chains diverge
…benchmarks; every CI pytest run deselects benchmarks explicitly
…h a merge commit, pynumerics version and changelog in the metadata checklist, and the PyPI publish and verification steps
…README and docs overviews for 2.3.0, and the pynumerics changelog and PyPI metadata
… copy as a safe git directory, so CI checkouts owned by the runner user pass git's ownership check
…aled ulps off Windows, where Exp and Log come from a different platform math library
@HadenSmith
HadenSmith merged commit f97fb00 into main Nov 10, 2026
7 checks passed
@HadenSmith
HadenSmith deleted the release/2.3.0 branch November 11, 2026 18:00
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant