0129 — Four .NET numerics libraries read, and three absence claims withdrawn

Status: accepted · Date: 2026-09-14 · Amends: 0105, 0116 · Applies: 0074, 0075

Context

#676 listed five .NET numerics libraries this repository had never read: NumFlat, Meta.Numerics, Numerics.NET, Extreme Optimization and ILNumerics. Decision 0074 says a gap is claimed on what a package exports, never on its README, and 0096 and 0105 each replaced one of this project's own claims by doing so. The issue asked for the same reading here, recorded even where the answer is uncomfortable.

It is uncomfortable. Three sentences this repository publishes are false as written.

The reading

Read on 2026-09-14 with tools/survey.cs (decision 0110); the member count is the surveyor's first line, accessors excluded. Every licence was read from the package, per decision 0075.

dotnet run tools/survey.cs -- Meta.Numerics 4.2.0 'PrincipalComponent|LjungBox|Autocovar'
dotnet run tools/survey.cs -- Numerics.NET 10.7.0 'PartialAutocorrelation|Kpss|VarianceInflation'
dotnet run tools/survey.cs -- ILNumerics.Toolboxes.MachineLearning 7.4.62 'pca$'
dotnet run tools/survey.cs -- NumFlat 1.3.4 'EigenValues|Regression'

	Meta.Numerics 4.2.0	Numerics.NET 10.7.0	ILNumerics 7.4.62	NumFlat 1.3.4
published	2025-07-14	2026-07-21	2026-08-25	2026-07-18
licence, from the package	MS-PL, <license type="expression">	commercial: License.md, an ExoAnalytics Inc. agreement, requireLicenseAcceptance	commercial: LICENSE.txt, ILNumerics GmbH's agreement, whose free plan excludes commercial use; requireLicenseAcceptance	MIT, <license type="expression">
lib/	netstandard2.0	net462, netstandard2.0, net8.0, net9.0, net10.0	net461, netstandard2.1, net8.0	net8.0
assembly read	Meta.Numerics.dll: 177 types, 1,644 members	Numerics.NET.dll: 905 types, 13,839 members	ILNumerics.Toolboxes.Statistics.dll: 513 types, 3,644 members; ILNumerics.Toolboxes.MachineLearning.dll: 312 types, 2,070 members	NumFlat.dll: 123 types, 938 members, as 0116 recorded

The five are four. Numerics.NET's own package description reads "formerly Extreme Optimization Numerical Libraries for .NET", and Extreme.Numerics stopped at 8.1.25 on 2024-05-28. And Meta.Numerics is not six years stale: the issue saw 4.1.4 of 2020-08-21, and 4.2.0 followed on 2025-07-14.

What each exports that this repository had called absent

subject	Meta.Numerics	Numerics.NET	ILNumerics	NumFlat
variance a principal component explains	PrincipalComponent.VarianceFraction, .CumulativeVarianceFraction	PrincipalComponent.ProportionOfVariance, .CumulativeProportionOfVariance	MachineLearning.pca(A, outWeights, outCenter, outScores)	PrincipalComponentAnalysis.EigenValues
autocorrelation	TimeSeries.Autocovariance	TimeSeriesFunctions.AutocorrelationFunction	none	none
partial autocorrelation	none	TimeSeriesFunctions.PartialAutocorrelationFunction	none	none
Ljung-Box	TimeSeries.LjungBoxTest	Tests.LjungBoxTest	none	none
stationarity	none	AugmentedDickeyFullerTest, KpssTest	none	none
hypothesis tests	Univariate.StudentTTest, MannWhitneyTest, KruskalWallisTest, KolmogorovSmirnovTest, OneWayAnovaTest, ShapiroFranciaTest; Bivariate.WilcoxonSignedRankTest; BinaryContingencyTableOperations.FisherExactTest; ContingencyTable<,>.PearsonChiSquaredTest	a Statistics.Tests namespace, Shapiro-Wilk, Anderson-Darling and Friedman among it	distributions and descriptive statistics, no test	none
regression inference	Parameter.Estimate as an UncertainValue (Uncertainty, ConfidenceInterval); GeneralLinearRegressionResult.F, .RSquared, .Anova. No adjusted R-squared, no VIF, and Parameter carries no test of its own	Parameter<T>.StandardError, .Statistic, .PValue, .GetConfidenceInterval; RegressionModel<T>.AdjustedRSquared, .FStatistic; LinearRegressionModel.VarianceInflationFactors	none	coefficients, intercept and prediction only

Searched for by name and found in none of the five assemblies above: Kaplan-Meier, Nelson-Aalen, the log-rank test, proportional hazards, censoring, conformal prediction, and mixed or random effects. The only Hazard members are Meta.Numerics' distribution hazard functions.

What is now false as written

1. "Below net8.0 nothing in .NET could say how many components a projection should keep." 0116 said the explained variance "has no incumbent on netstandard2.0", and docs/reference/decomposition/factorization.md, docs/migration/sklearn.md, README.md's incumbent table and both benchmark pages repeated it. Meta.Numerics reports it on netstandard2.0 under MS-PL, and Numerics.NET does under a commercial licence. 0116 read the two PCA incumbents a reader was being pointed at, and the sentence reached further than the reading.
2. "No maintained .NET package carries" the time-series diagnostics. 0105 said "what nobody maintained ships is the apparatus for deciding whether a series may be modelled at all", and docs/migration/statsmodels.md said the same. Numerics.NET ships the autocorrelation and partial autocorrelation functions, Ljung-Box, ADF and KPSS, and published this July. Meta.Numerics ships the autocovariance and Ljung-Box.
3. Accord.Statistics is "the one .NET library that did carry" hypothesis tests, and there is "nothing maintained to benchmark against" — docs/guides/hypothesis-testing.md. Meta.Numerics carries eight of the ten families Lodestar.Stats ships, on netstandard2.0, and Numerics.NET more.

A fourth is false in a smaller way. README.md says the standard errors, t and p values, intervals, adjusted R-squared, F test and VIF of a regression "are in none of them". Numerics.NET exports every one; Meta.Numerics exports the standard error, the interval and the F test.

What still holds

• Survival and conformal prediction have no .NET incumbent, commercial ones included. README.md's items 4 and 7 and decision 0099 stand.
• The free inference table is still incomplete. Meta.Numerics stops before the adjusted R-squared, the VIF and a test per coefficient, which is what 0096 called the gap: "a maintained, permissively licensed, framework-free OLS table". NumFlat returns coefficients only.
• A free partial autocorrelation, stationarity test or seasonal decomposition exists only in Cortex.TimeSeries, which 0105 already recorded and set aside.
• NumFlat changes nothing 0116 did not already record. Its PCA is measured (#701), its regression has no inference, and its four clustering algorithms against Lodestar.Cluster's one are #681's measurement. That answers #676's last criterion: NumFlat is already a measured incumbent for Decomposition, and #681 makes it one for Cluster.

Decision

Nothing already shipped is withdrawn, and every sentence above is corrected where it was written. 0105 and 0116 are amended on their absence claims; their decisions stand.

• A commercial incumbent is named, not delegated to. A migration row recommends what a reader can install without buying a licence, as 0105 did for Dew.Stats. Numerics.NET is now named beside each subject it covers, with its licence, so a reader who already holds one knows it answers the question.
• Meta.Numerics is named as a free incumbent and delegated to for nothing yet. PrincipalComponentVariance, SerialCorrelation and the Lodestar.Stats families stay. What they add over it is checkable rather than asserted: agreement with scikit-learn, statsmodels and scipy at 1e-9 through a frozen corpus, confidence bands and a partial autocorrelation it does not export, and an Apache-2.0 span API. What they do not add is availability below net8.0, and no document says so any more.
• Meta.Numerics is measured before any number is claimed against it: #756 adds it to bench/ for the explained variance and the hypothesis tests it shares with Lodestar.Stats.

Options that lost

• Deprecate PrincipalComponentVariance in favour of Meta.Numerics. Had this reading preceded 0119, 0116's reopening condition might have been met by a migration row. It did not, and the type shipped in Lodestar.Decomposition 0.3.0 with a frozen scikit-learn corpus that Meta.Numerics has never been held to. Retiring it on a member name, before #756 has checked that the two agree, would repeat the mistake this record corrects in the opposite direction.
• Correct only the ADR. #676's third criterion refuses it, and so does the reason 0074 exists: a reader meets the claim in the migration page and the README, not in a decision record.
• Leave the commercial libraries unnamed, since nothing is delegated to them. An absence claim that is true only once commercial software is excluded has to say so, or it is false.

Consequences

• README.md's item 6 and its Lodestar.Decomposition and Lodestar.Stats incumbent rows name what these libraries export.
• The PCA pages — docs/migration/sklearn.md, docs/reference/decomposition/factorization.md and its PrincipalComponentVariance page, docs/equivalence.md's projection row, bench/README.md section 30 and docs/guides/performance.md — name Meta.Numerics beside NumFlat and ML.NET. docs/guides/hypothesis-testing.md lists the three libraries that carry tests, and docs/migration/statsmodels.md's hypothesis-test row stops pointing at Math.NET, which ships none.
• docs/migration/statsmodels.md's time-series row splits and stops reading being written: #617 shipped the serial-correlation half, and #671 holds the rest, where Numerics.NET's ADF and KPSS are now on record.
• #621 inherits one line of evidence: none of these four exports a mixed or random-effects member.