各バージョンの新機能

v7.0での新機能 (2018/12)

.NET Core and .NET Standard support

• Support for .NET Core 1.1 and 2.1.
• Support for .NET Standard 1.3 and 2.0.
• Support for .NET Framework 3.5, 4.0, 4.72 and later.
• All packages are available on the Nuget Gallery.

Linear algebra

Major enhancements

• Broadcasting vectors in matrix operations.
• Enable Conditional Numerical Reproducibility option for native libraries.
• Upgraded native libraries to Intel® Math Kernel Library version 2019 Update 0.
• Upgraded managed linear algebra library to LAPACK 3.7.0.
• Improved range and accuracy of matrix exponential.
• Vector Map methods that include index as delegate argument.

New matrix decompositions

• Generalized Eigenvalue Decomposition.
• Generalized Singular Value Decomposition (GSVD).
• Sparse singular value decomposition.
• RQ decomposition, QL decomposition, and LQ decomposition.
• Access to 'thin' version of the orthogonal factor Q in a QR decomposition.
• Compute factors of symmetric and Hermitian indefinite decomposition.

Performance improvements

• Improve performance for level 2 managed sparse BLAS.
• Improve performance for various vector operations.
• The threshold for parallel execution of vector maps can now be configured.

Mathematics

General improvements

• The generic Operations<T> class has been optimized to eliminate nearly all overhead for the most frequently used operations on the most common argument types.
• ParallelOptions is now exposed for all algorithms to enable cancellation and other scenarios.
• Combinatorial iterators to enumerate all combinations, permutations, and Cartesian products of sets of items.
• New overloads for numerical integration methods that take Interval objects to specify bounds.
• Inverse hyperbolic functions for decimal and quad precision numbers.

Optimization

• The NonlinearProgram class has a new constructor that accepts variable names.
• Symbolic constraints that are linear in the variables are now recognized as such.
• The Nonlinear Program solver can now recover when it encounters an infeasible subproblem.
• Up to 30% improvement in the performance of the Linear Program solver
• Limited Memory BFGS Optimizer.
• LeastSquaresOptimizer base class for nonlinear least squares algorithms.
• Trust Region Reflexive algorithm for nonlinear least squares.
• Trust Region Reflexive algorithm option in nonlinear curve fitting.
• Improved documentation for nonlinear least squares algorithms.

Special functions

• Jacobi elliptic functions.
• Zeros of Bessel and Airy functions.
• The performance and accuracy of Bessel functions of the first and second kind has been improved.
• Polygamma function.
• Modified Bessel functions of real order.
• "Partial application" functions for incomplete and regularized Gamma and Beta functions.
• Zernike polynomials.

Statistics and data analysis

Data access library

• Data Access Library providing a unified API for reading and writing data frames, matrices, and vectors.
• Reading and writing R's .rda/.rdata and .rds files.
• JSON serialization.
• Other supported formats include: delimited text (CSV, TSV...), fixed-width text, Matrix Market, Matlab®, stata®

Statistical models

• Use R-style model formulas to specify statistical models.
• Partial Least Squares (PLS) models.
• Linear Discriminant Analysis.
• Kernel Density Estimation.
• Binomial Generalized Linear Model can now be used with count data.
• Two-way ANOVA: support for Type I, Type II, and Type III sums of squares.
• New ConditionalVariances property on GARCH models.
• The performance of ARIMA model fitting has been improved.
• Nicer Summarize for statistical models.

Hypothesis tests

• Augmented Dickey-Fuller test.
• Cramer-von Mises Goodness-of-fit test.
• Tests for outliers: Grubbs' test, Generalized ESD test.

Data analysis

• New aggregators: Range, Mode, CountUnique.
• Improved support for custom aggregators based on accumulators.
• R-style variations of quantiles.
• LOESS and LOWESS smoothing.
• More categorical encodings: Backward difference, Forward difference, Helmert, reverse Helmert, orthogonal polynomial encoding.
• Non-central chi-square, non-central F, non-central beta, and non-central t distributions.
• Anderson Darling distribution is now public.

v5.1での新機能 (2014/11)

• Data frame class supporting advanced data manipulation.
• Support for GPU computing
• Offload computations to the GPU.
• Support for quadruple precision numbers with up to 34 digits, including our full vector and matrix library. Real and complex numbers are supported.
• Smoothing cubic splines, monotonicity preserving splines.
• GARCH models (Generalized AutoRegressive Conditional Heteroskedasticity).
• Regularized linear regression models: Ridge regression, LASSO and elastic net.

v4.2での新機能 (2011/11)

• Automatic differentiation: symbolic computation of derivatives, gradients and Jacobians.
• Extensible with built-in support for derivatives of methods in System.Math and most elementary and special functions in the library.
• Backward differentiation with common sub-expression elimination generates optimal evaluation.
• New SymbolicMath class that lets you optimize functions and solve equations specified as lambda expressions using automatic differentation.
• Evaluation of (sequences of) classic orthogonal polynomials: Chebyshev (1st and 2nd kind), Hermite, Laguerre, Legendre and Gegenbauer.
• Stepwise linear regression.
• Regression fits of linearized curves: logarithmic, power, exponential, reciprocal...
• 2x2 and RxC Contingency tables.

v4.0での新機能 (2010/12)

• .NET Framework Version 4.0 とVisual Studio 2010の並列処理サポート
• 複数コア利用による計算スピードの向上 (.NET 4.0のみ)
• 新しい疎線形プログラムソルバーは、100万変数の以上の問題を解くことが可能
• ブランチとバウンドアルゴリズムを使った混合整数線形計画法
• 新しい特殊関数: ハイパー幾何, Riemann zeta, 楕円積分, Fresnel関数, Dawsonの積分
• FFTのウィンドウ関数の完全セット
• 線形代数および任意精度のタイプでの6xまでのパフォーマンスの向上
• F#での新しい50個のクィックスタートサンプル