# Changelog All notable changes to this project will be documented in this file. The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). ## [1.10.0] - 2026-08-19 ### Added - Added `TVector2D.Rotate(const Angle: Double): TVector2D` to `GeometryLib.Geometry` as an allocation-free value operation. `Angle` is in radians; positive angles rotate counter-clockwise about the origin; the source vector is not modified. `Rotate(Pi / 2)` agrees with `Perpendicular` and `Rotate(-a)` inverts `Rotate(a)` within a ~1e-15 relative tolerance on the tested ordinary and extreme finite ranges; no universal relative bound is claimed for every finite or denormal `Double`; the zero vector rotates to the exact zero vector; non-finite angles or components follow the documented IEEE-754 convention. - Implemented the closed [1.10.0 capability manifest](docs/CAPABILITY_MANIFEST_1.10.0.md) in full. The closed deprecation-marking decision is no-deprecation: no declaration is deprecated, removed, or moved, and full 1.x source compatibility is retained. ### Documentation - Documented the `TVector2D.Rotate` contract in the [GeometryLib reference](docs/GeometryLib.md) and extended the runnable [geometry example](examples/12_geometry.pas) with a rotation workflow whose output is now checked by the example-output gate. - Regenerated the [API reference](docs/API_REFERENCE_1.9.md) and snapshot to record the single `TVector2D.Rotate` addition, and recorded the now-decided 2-D rotation contract in the 2.0-candidate decision and diff. - Published the [1.10.0 release notes](docs/RELEASE_NOTES_1.10.0.md). ## [1.9.9] - 2026-08-17 ### Added - Published the closed 1.10.0 capability manifest: `TVector2D.Rotate` is declared for 1.10.0 with a complete behavior contract, test plan, documentation plan, and compatibility impact; the deprecation-marking decision closes as no-deprecation; every other proposal is explicitly deferred beyond 2.0 with a recorded reason. - Added an offline convergence checker that validates the manifest against the roadmap, the 1.9.3 API decision, the 1.9.7 migration rehearsal, the frozen 1.9 API snapshot, and the capability inventory, and fails when any API design question remains open. - Published the final 1.9.9 candidate API snapshot and exact diff from 1.9.0, separating compatibility-preserving corrections (none), 2.0 documentation changes, and packaging changes (none). ### Governance - Completed the algorithm/fixture provenance and licence audit covering every stable `src/` unit, the 2.x maintenance policy, support policy, deprecation policy, contribution gate for new domains, and the security support window. ### Validation - Integrated the convergence gate into ordinary CI, clean-archive release qualification, and the release checklist. ### Documentation - Published the capability manifest, final snapshot and diff, provenance audit, governance policies, release notes, PR notes, and qualification record for 1.9.9. ## [1.9.8] - 2026-08-16 ### Added - Added three maintained multi-domain end-to-end workflows: a sensor pipeline (DSP + statistics + time-series), a numerical modelling and optimisation workflow (fitting + interpolation + solve + bounded optimisation), and a reproducible probability/finance analysis workflow (seeded sampling + inference + NPV/IRR). Each loads or seeds local data, validates it, transforms, fits or solves, runs a diagnostic path, and exports a deterministic report. - Added an offline workflow-qualification contract and checker that validates the machine-readable manifest, compiles and runs every workflow with FPC 3.2.2 from an isolated work directory, verifies success markers, diagnostics, numerical bounds, exported artifacts, and byte-identical repeatability, and records the exact compiler/platform that ran. ### Validation - Integrated the workflow checker into ordinary CI and clean-archive release qualification so representative workflows are reproduced from published instructions without network access, foreign runtimes, licence keys, or repository-local configuration. ### Documentation - Published the workflow guide, release notes, PR notes, qualification record, and updated support/capability/release identity for 1.9.8. ## [1.9.7] - 2026-08-12 ### Added - Added independent supported-1.x and candidate-2.0 consumer projects that compile and assert construction, ordinary results, diagnostic failure, ownership/copying, indexing, precision, defaults, and interpretation across all 13 documented domains. - Added an offline migration-rehearsal contract and checker covering direct FPC use, the Lazarus package boundary, expected warnings/source edits, and generated host-specific results. - Published conservative NumLib and LMath/DMath conceptual mappings with explicit storage, indexing, ownership, diagnostic, precision, licensing, and unsupported-compatibility differences. ### Compatibility - Retained `TPressureKit`, `EPressureError`, `TVelocityKit`, and `EVelocityError` in their existing units after executable source/package rehearsal confirmed exact identity and no hidden dependency. No deprecation, warning, package move, or removal runway is introduced. ### Documentation - Published the migration decision list, side-by-side source edits, package evidence, release notes, PR notes, and qualification record for 1.9.7. ## [1.9.6] - 2026-08-11 ### Validation - Added a native Pascal ABI, endian, locale, numerical, and exact binary-format probe plus an offline validator for tiered target evidence and stable-source/ package dependency assumptions. - Release qualification now verifies source-archive SHA-256 and exact clean extraction, rejects repository/compiler artifacts, extracts and rechecks the offline documentation ZIP, and retains target-specific portability results. - Primary checks run on each change and complete Linux/Windows archive qualification runs weekly and for exact candidates/releases; Windows i386 has a bounded secondary profile rather than inferred primary coverage. - Clean archive workflows block new outbound connections after toolchain installation and restore runner networking before evidence publication. ### Documentation - Published compiler, OS/CPU, pointer/scalar ABI, evidence dates/refs, exact checks, platform-assumption audit outcomes, offline installation steps, and visible unqualified-target limits in machine-readable and human-readable v1.9.6 evidence. ## [1.9.5] - 2026-08-11 ### Performance - Reused the tested native radix-2 FFT kernel inside applied DSP transforms, removing repeated per-butterfly trigonometric evaluation while retaining a readable portable fallback and the existing public contract. - Added representative cold and warmed benchmark rows for dense, sparse, iterative, DSP, modelling, statistics, and data-analysis workloads at small and large scales. ### Validation - Added exact allocation, retained-state, dense-shape, complexity, and checksum gates plus advisory same-run and host-matched prior timing comparisons. - Integrated the checked performance evidence into offline release qualification; variable-runner wall-clock movements request review instead of creating false hard failures. ### Documentation - Published the benchmark conditions, scalar kinds, dimensions, setup rules, tolerances, compiler flags, host metadata, claim mappings, and limitations in machine-readable and human-readable v1.9.5 evidence. ## [1.9.4] - 2026-08-10 ### Documentation - Published a machine-readable numerical-evidence catalogue for every stable capability family, with bounded domains, edge cases, evidence categories, budgets, test links, and reproducible reference provenance. - Added a human-readable evidence report that distinguishes exact properties, reference comparisons, residual/reconstruction checks, and feasibility diagnostics without claiming universal numerical proofs. ### Validation - Added offline catalogue validation to reject missing or malformed coverage, invalid numerical budgets, absent cited sources/tests, and inconsistent high-risk fault-injection records. - Added isolated compile-and-test mutation checks for scalar special functions, typed dense direct solving, and Bluestein DSP transforms; each sampled fault must be detected by FPCUnit. ## [1.9.3] - 2026-08-04 ### Documentation - Completed the proposed 2.0 API decision by classifying every exact 1.9 declaration, publishing a 13-domain common-path map, and resolving shared naming, shape, ownership, error, tolerance, outcome, RNG, callback, and thread-safety conventions. - Gave every compatibility declaration a named typed replacement with semantic differences or an explicit retain decision, and published an exact empty source/behavior/warning/packaging diff separate from documentary defaults. - Reviewed all 21 plain compiler aliases for identical behavior, defaults, ownership, exception identity, and numerical results. Retained 17 aliases and routed the four pressure/velocity facade and error aliases to 1.9.7 migration and package-boundary confirmation without deprecating them. - Routed ergonomic 2-D and 3-D vector rotation to focused 1.10.0 design work without adding an unreviewed patch-release declaration. ### Validation - Added exact decision, selector/overload, generic-owner, domain/unit, alias-review, compatibility, common-example, convention, and proposed-diff checks to the release qualification suite. - Regenerated the exhaustive 2,880-row API reference with five-way classifications and compatibility decision/note columns while preserving all frozen 1.9 interface hashes. ## [1.9.2] - 2026-08-02 ### Documentation - Added a double-real beginner guide, task-oriented recipe index, and explicit beginner-to-advanced routes for every stable domain without changing the frozen 1.9 public API. - Added searchable routes for dense and sparse solves, descriptive and streaming statistics, probability, interpolation and fitting, optimisation, FFT and filtering, time series, finance, geometry, and unit conversion. ### Validation - Made each beginner recipe's code and claimed output part of the clean- archive documentation checks, including all 13 domain routes, generated problem-oriented search results, local links, release identity, and exact public declarations. - Made checked-heap qualification capture heaptrc in an explicit cross-platform log, reject missing or nonzero summaries numerically, and preserve evidence artifacts even when a qualification job fails. ## [1.9.1] - 2026-08-02 ### Fixed - Replaced seeded bootstrap's private low-bit LCG/modulo sampling with the shared explicit-state generator's unbiased bounded-index path. Eight-element samples no longer collapse every resample to the same mean; a permanent regression covers varying resamples, non-degenerate confidence bounds, deterministic replay, and unchanged global random state. ### Documentation - Added release-identified versioned static documentation, preservation of the tagged 1.9.0 site, and a deterministic offline HTML ZIP plus SHA-256 file. - Made every output-producing runnable documentation program verify its claimed output, and added checked statuses/final markers for the release- facing dense, sparse, and migration examples. - Added a focused feedback route for installation time, type choice, conversions, unexpected errors/statuses, selection guidance, and migration. ### Validation - Added clean-archive qualification automation for normal, optimised, checked/heap-traced, examples, documentation, Lazarus package, quick-start, and benchmark gates, while preserving the exact 1.9 public API snapshot. - Recorded the release evidence and unchanged limitations in `docs/QUALIFICATION_1.9.1.md`. ## [1.9.0] - 2026-08-01 ### Added - Immutable validated CSR/CSC and diagonal/tridiagonal/band storage for single/double real and complex values, deterministic triplet construction, sparse arithmetic/products, and explicit dense conversions. - Typed retained/delegated linear operators and identity, diagonal, IC(0), and ILU(0) preconditioners with documented ownership and reentrancy. - CG, MINRES, restarted GMRES, BiCGSTAB, and LSQR with shared options, true-residual diagnostics, cancellation/progress, breakdown reasons, and reusable scalar-specific workspaces. - Reusable pivoted tridiagonal/no-pivot band factors and an explicit natural-order sparse LU baseline with multiple RHS and fill diagnostics. - Deterministic restarted largest-magnitude Lanczos and Arnoldi with independently checked residuals. - Matrix Market coordinate double-real/double-complex sparse exchange and versioned checksummed four-scalar sparse binary interchange. - End-to-end sparse and candidate-2.0 migration examples, a classified machine-readable public-API snapshot, declaration/default documentation checks, capability records, and large no-densification benchmarks. ### Validation - Added dense-oracle, canonical/malformed interchange, status/breakdown, factor/preconditioner/workspace reuse, in-place alias, concurrent immutable reuse, and 20,000-dimensional no-densification fixtures. - Corrected matrix-free shape validation to bound vector axes independently, preserving large linear-storage workloads on Win32 without weakening invalid-dimension or address-space checks. - Recorded 100,000-nonzero sparse and 200,000-dimensional matrix-free `-O3` qualification cases with nonzero/allocation/storage/product/iteration/ residual/timing evidence. - The complete platform, package, example, documentation, and archive matrix is recorded in `docs/QUALIFICATION_1.9.0.md`. ### Compatibility - Existing public signatures, defaults, and `TMatrixKitSparse` behavior remain available. Legacy-to-typed sparse conversion is explicit. - The 2.0 contract and migration example are preview evidence only. No 2.0 removal, default change, or later-roadmap implementation is included. ## [1.8.0] - 2026-07-30 ### Added - Explicit local xoshiro256** random state and constant-memory mergeable online statistics, with replay, split-stream, and bounded-state tests. - Arbitrary-length/single/batched/2-D FFTs, direct/FFT/overlap convolution, correlation, rational resampling, spectra, analytic signals, Haar transform, and bounded block/FIR/biquad processing. - Paired distribution APIs, parameter estimates, common inference/effect/ correction results, SVD OLS diagnostics, and separation-aware binary logistic regression. - Typed-dense PCA, seeded k-means++, hierarchy, fitted standardization, deterministic splits, binary LDA, seeded classification/regression forests, an exact low-dimensional KD tree, and scalar/multivariate Kalman filtering. - Invariant text, delimited text, a dense Matrix Market subset, and versioned checksummed little-endian numerical and random-state persistence. - Versioned selected-model adapters, typed metadata/complex summaries, and a resource-bounded scalar/vector/matrix expression evaluator. - Audited 1.7 gap closure: classical spline boundaries, explicit complex-step and vector AD, cubature/local-RNG Monte Carlo, scaled/covariance-aware fitting, complex polynomial roots, component ODE tolerances, detailed optimization/workspaces, two-phase simplex, and QP outcome certificates. - Deterministic serial blocked and automatic typed dense multiplication using the portable kernels as cross-path correctness oracles. - Cross-domain and interchange/replay examples, capability records, release notes, benchmark comparison, and qualification evidence. ### Fixed - Made simultaneous complex polynomial-root initialization non-conjugate-symmetric, removing a platform-rounding-dependent Win32 convergence failure. ### Validation - Release-qualified on 2026-07-30 with 899 passing tests under Win64 normal, `-O3`, and checked-heap builds, plus 899 passing tests under Win32 `-O2`. - Built and ran all 22 examples, built both Lazarus packages for Win64 and Win32, and passed documentation checks covering 50 pages, 22 examples, and 248 public symbols. - Rebuilt and verified the checksummed clean source archive, then reran its 899-test suite, all 22 examples, and documentation checks. - Passed the required Linux and Windows GitHub Actions pull-request and push workflows on the qualified release branch. ### Compatibility - Existing 1.7 and earlier public signatures remain available. The 1.8 APIs are additive; the legacy `EngineeringLib.Signal` FFT remains unchanged. - No parallel/SIMD ABI, advanced filter-design/broader-wavelet surface, survival/factor/robust-covariance family, implicit stiff ODE, or general model/decomposition persistence is claimed stable. ## [1.7.0] - 2026-07-30 ### Added - Reusable barycentric, rational, PCHIP, Akima, bilinear/bicubic, IDW, RBF, and thin-plate interpolation with explicit ownership and scale limitations. - Scale-aware finite gradients/Jacobians/Hessians, forward dual-number automatic differentiation, and pre-solve derivative checking. - Adaptive Gauss-Kronrod finite/improper integration, deterministic quasi-Monte-Carlo integration, typed-QR linear fitting, damped robust nonlinear least squares, vector Newton equations, and adaptive Dormand-Prince ODEs with dense output and event detection. - A shared `TIterationStatus` diagnostic vocabulary and dense convex QP/SOCP workflows with objective, feasibility, optimality, iteration, and evaluation results. - End-to-end modelling and convex-optimisation examples, selection guides, tests, capability records, release notes, and qualification evidence. ### Changed - `PenaltyMethod` and `Maximize` no longer serialize through unit-global callback bridges; their callback paths are reentrant. - The Lazarus package and distribution metadata now identify version 1.7.0 and include every new unit. ### Compatibility - Existing 1.6 and earlier public signatures remain available. The 1.7 APIs are additive. ## [1.6.0] - 2026-07-27 ### Added - Added reusable single/double real/complex Householder QR and column-pivoted QR factors, full-rank and rank-revealing least-squares solves, copied compact factors/permutations, numerical-rank decisions, and inspectable residual/backward-error diagnostics. - Added reusable compact one-sided Jacobi SVD factors for tall, square, and wide matrices plus vector/multiple-RHS rank-deficient and underdetermined minimum-norm solves. - Added full ascending symmetric real and Hermitian complex eigensystems with normalized eigenvectors, deterministic cyclic-Jacobi convergence outcomes, and single/double scalar parity. - Added lower/upper, unit/non-unit triangular solves for ordinary, transposed, and conjugate-transposed systems. - Extended LU and Cholesky factors additively with `ConditionIndicator` and `SolveWithInfo`, and added diagnostic square and positive-definite convenience entry points without changing the 1.5 solve contract. - Added the 1.6 design record, solver-selection and migration guides, realistic calibration/minimum-norm/eigen example, capability metadata, focused public/numerical tests, deterministic factor-reuse benchmarks, release notes, and qualification report. ### Compatibility - Preserved every `IMatrix`, `TMatrixKit`, `IVector`, and typed 1.5 entry point. No API was removed or deprecated, and compatibility decompositions are not silently rerouted where their contracts differ. ## [1.5.0] - 2026-07-26 ### Added - Added 32-byte-aligned contiguous row-major single/double real/complex dense matrices with checked native-size shapes, retained-owner mutable views, explicit clones, and explicit compatibility conversions. - Added matching allocating and `Into` dense kernels for elementwise arithmetic, AXPY, scalar-function application, reductions, transpose/conjugation, compensated products, and dot products. - Added allocation-free operator-friendly 2x2 real/complex value matrices, batch types, and explicit flat/nested, precision, real/complex, and compatibility conversions. - Added pivoted-LU `Solve(A, B)`, reusable LU factors, and real/complex Cholesky factors for vector and multiple right-hand sides. - Added `TSingleComplex`, the typed dense migration/design guides, capability inventory, support matrix, solve example, benchmark, and completion-gate tests. ### Compatibility - Preserved `IMatrix`, `TMatrixKit`, `IVector`, and all existing unit entry points. No API was removed or deprecated. ## [1.4.0] - 2026-07-25 ### Added - Added componentwise `+`, binary and unary `-`, scalar `*` in both operand orders, and vector/scalar `/` operators for `GeometryLib.Geometry` `TVector2D` and `TVector3D`. The fixed-size value records retain IEEE-754 `Double` results for signed zero, NaN, infinity, overflow, and zero-scalar division without mutating either operand. - Added focused GeometryLib operator, edge-case, algebraic-property, and public-API smoke coverage, plus a runnable Theodorus-spiral demonstration in the geometry example. ### Fixed - Made `TVector2D` and `TVector3D` magnitude and normalization scale-safe for finite tiny and large components. Normalization now accepts every finite non-zero vector and rejects exact-zero or non-finite vectors explicitly. ## [1.3.0] - 2026-07-23 ### Added - Added the 1.3.0 complex and vector foundation with `MathBase.Complex`: the double-precision `TComplex` record, stable arithmetic and magnitude, principal elementary functions, and `TComplexArray`. - Added `AlgebraLib.VectorKernels` and its `TVectorKit` facade for finite, contiguous real and complex array-vector arithmetic, including scale-safe norms and conjugating complex dot products. `AlgebraLib.Vectors` continues to preserve the existing matrix-as-vector aliases while re-exporting the new array-vector API. - Added `TComplexArray` FFT/IFFT overloads that preserve the existing split-real/imaginary signal API, documentation, a runnable complex-vector example, and focused tests. - Hardened complex division and principal branch behavior, added inverse complex trigonometric and hyperbolic functions, and made `TComplexArray` the native FFT representation. - Expanded vector kernels with stable reductions, elementwise arithmetic, and allocation-avoiding `...Into` destination-buffer procedures. - Added representative complex, vector, and FFT benchmarks plus public API smoke coverage. ### Fixed - Preserved tiny `CAsinh`/`CAtanh` inputs and replaced target-sensitive `z*z` cancellation with scaled component formulas for large inverse complex functions. - Defined infinity/NaN behavior for complex magnitude, exponential, and square root calculations, and retained signed-zero branch sides for inverse hyperbolic functions. - Updated CI to compile and execute the standalone `.pas` examples after their rename from `.lpr`. ## [1.2.3] - 2026-07-21 ### Changed - Reworked `MathBase.Precision` around a higher-accuracy Lanczos log-gamma, cancellation-resistant beta evaluation, convergence-checked incomplete-beta fractions, and direct incomplete-gamma normal tails. - Normal, lognormal, beta, Student-t, and F survival functions now evaluate upper tails directly instead of subtracting a rounded CDF; F-distribution arguments avoid overflowing `d1*x`. Normal-tail approximations in statistics use the same direct-tail path. - Hyperbolic and inverse-hyperbolic helpers preserve tiny arguments and avoid unstable exponential/logarithmic cancellation. Two-dimensional vector magnitude now uses the scaled hypotenuse kernel. ### Fixed - Corrected `MathBase.Precision.StudentT` to use the required `df/2` incomplete-beta shape parameter instead of `(df+1)/2`, fixing t-test p-values built on the shared helper. - Invalid low-level special-function shape parameters now return NaN predictably, and iterative special functions no longer return an unconverged partial result. - Rebased the K-S normality reference on the full-precision normal CDF instead of the previous low-accuracy error-function approximation, and made its empirical CDF fractions explicitly double precision across FPC targets. ### Tests - Added special-function reference, symmetry, invalid-input, small-argument, large-vector, and representable extreme-tail regressions, bringing the suite to 798 tests. Cross-platform cases now pin complement inputs and empirical fractions to `Double` before comparing extreme-tail and K-S references. ## [1.2.2] - 2026-07-18 ### Documentation - Added newcomer-oriented MathBase and NumericsLib walkthroughs, giving every documented domain a runnable example, plus an indexed learning path for all 14 examples. - Reviewed all examples for concise purpose, input, interpretation, and method- selection guidance; corrected the statistics walkthrough to use a local, reproducible bootstrap seed and clarify p-value interpretation. - Added 1.2.2 release notes and updated version and release references. ### Tooling - Added shell and PowerShell entry points that compile every example into an ignored `example-bin/` directory, and made CI exercise both the scripts and the resulting programs on their native platforms. ## [1.2.1] - 2026-07-18 ### Documentation - Defined consistent terminology for the mathlib-fp project, domains, Pascal unit families, units, Kit classes, focused aliases, and the Lazarus package. - Added a public API naming inventory and aligned guides, source headers, and contributor guidance with it without renaming existing public identifiers. ### Tests - Added compile-time smoke coverage for every documented Kit class and focused alias, and registered the existing Engineering focused-alias tests in the main test runner, bringing the current suite to 789 tests. ## [1.2.0] - 2026-07-18 ### Added - `EngineeringLib.Common` with `EEngineeringError` and domain-specific exceptions for fluid dynamics, thermodynamics, signals, and unit conversion. - Seeded `CreateRandom`, `BootstrapMean`, and `BootstrapConfidenceInterval` overloads that are reproducible without changing global random state. - `PolynomialFeatures(..., IncludeBias)` overload so callers can omit the bias column when fitting models that already estimate an intercept. - Edge, property, residual, deterministic-randomness, FinanceLib focused-unit, rounding, UTF-8, parallel-multiplication, and numerical reference coverage, bringing the suite to 788 tests. - Representative performance benchmarks for statistics sorting, convex hulls, and dense matrix multiplication, with CI compilation coverage. - Focused `EngineeringLib.Velocity` and `EngineeringLib.Pressure` entry units now expose directly nameable exception aliases and have direct compilation and runtime coverage. ### Changed - Replaced hard-coded eigendecomposition and power-method cases with general real algorithms and residual-based convergence checks. - Fractional matrix powers now use the symmetric eigendecomposition and require a symmetric positive-definite matrix; integer powers use exponentiation by squaring. - Matrix multiplication uses an operation-count threshold, caps worker count, propagates worker failures, and falls back safely on Unix programs without a thread manager. - ML entry points consistently reject empty, ragged, non-finite, mismatched, or out-of-range inputs with `EMLError`. - Financial methods that expose `ADecimals` now apply it consistently. NPV rounds only its final result, and amortization schedules use the requested precision for payment amounts. - Undefined financial ratios now raise `EFinanceError` when a required denominator is zero instead of returning a fabricated zero. - `FinanceLib.Bonds` and `FinanceLib.NPV` remain lightweight focused entry units and now export directly nameable supporting aliases for their cash-flow and amortization types. - Random-producing library functions no longer call `Randomize` internally. - The test runner installs `cthreads` first on Unix and verbose algebra-test output is opt-in through `MATHLIB_TEST_VERBOSE`. - The Lazarus package version is now 1.2.0 and includes the shared engineering exception unit. - The Lazarus package and registration unit are now named `mathlib_fp` to match the mathlib-fp project name. - Lazarus 4.8 is the minimum supported Lazarus version; CI validates the package against that baseline. - `EngineeringLib.Signal.TDoubleArray` now aliases the shared `MathBase.SharedTypes.TDoubleArray` type. - `TFluidDynamicsKit.PumpHead` now accepts explicit inlet and outlet velocities and implements the Bernoulli velocity-head term `(v2²-v1²)/(2g)`. - General-purpose statistics and geometry sorts now use O(n log n) algorithms instead of quadratic insertion sorts. - Root finders expose detailed convergence records, iterative matrix and scalar solvers report exhaustion explicitly, PCA records per-component iterations, and linear programming exposes a precise termination status. - Linear regression now uses centered Householder QR instead of normal equations, improving behavior for high-offset and ill-scaled data. ### Fixed - Fixed example 11's unsupported format specifier and duplicate polynomial intercept, which previously caused a singular regression system. - Unknown unit names no longer silently default to length. Non-`Try` APIs raise `EUnitConversionError`; `Try...` APIs retain their `False` contract. - Significant-digit formatting now uses stable round-half-to-even behavior on both Win32 extended-precision and Win64 targets. - Fixed `InternalRateOfReturn`, which previously returned its initial 10% guess without iterating. It now brackets and bisects positive or negative rates and reports invalid or unbracketable cash-flow inputs with `EFinanceError`. - Corrected FinanceLib signatures, result-type scope, numeric examples, and exception contracts in the API guide and source comments. - Corrected odd-order high-pass and band-stop FIR centre indexing, added safe FFT/IFFT empty and mismatched-array handling, and documented the complete N-bin spectral outputs. - Humidity-ratio and adiabatic calculations now reject invalid pressure and specific-heat-ratio domains with `EThermodynamicsError`. - Engineering comments and API documentation now cover focused aliases, formula domains, signal shapes, every UnitConversion public API, exact unit names, fixed-duration time conventions, and locale-sensitive parsing. - Removed all compiler warnings from clean normal and UTF-8 builds. - Fixed broken source links, stale API names, version text, and random/bootstrap contracts across the documentation. - Corrected matrix inverse permutation handling, LU row swaps, scale-relative rank/singularity decisions, exact triangular cleanup, and large-norm matrix exponentials, including architecture-independent overflow reporting. - Corrected forward ray-circle semantics, polygon-boundary classification, non-negative radius validation, and zero-vector angle handling. - Corrected exact small-sample Mann-Whitney p-values, K-S D/p-value semantics, Shapiro-Wilk normal scores, and pooled-variance Cohen's d. - Corrected FFT period-bin mapping and ARIMA MA/integration forecasts. - Added checked combinatorics overflow paths and overflow-safe modular exponentiation, a Win32-safe sieve index path, plus stable hyperbolic and hypotenuse calculations. - Strengthened finite-value, dimension, domain, and callback validation across numerical, optimization, time-series, matrix, geometry, and ML entry points. --- ## [1.1.0] - 2026-04-16 ### Added #### NumericsLib — new library (`src/NumericsLib.Numerics.pas`, class `TNumericsKit`) A complete numerical methods library with no external dependencies. ##### Root Finding - `Bisection` — guaranteed bracketed convergence; raises `EInvalidArgument` when f(a)·f(b) ≥ 0 - `NewtonRaphson` — fast quadratic convergence using the function and its derivative - `Brent` — hybrid method combining bisection, secant, and inverse-quadratic interpolation; the recommended general-purpose solver - `Secant` — derivative-free quasi-Newton method requiring two initial guesses ##### Numerical Integration (Quadrature) - `TrapezoidalRule` — composite trapezoidal rule; O(h²) accuracy - `SimpsonRule` — composite Simpson's rule; O(h⁴) accuracy; auto-increments odd N to even - `GaussLegendre5` — 5-point Gauss-Legendre quadrature; exact for polynomials up to degree 9 ##### ODE Solvers (dy/dt = f(t, y)) - `EulerStep` / `EulerSolve` — 1st-order explicit Euler method - `RK4Step` / `RK4Solve` — classic 4th-order Runge-Kutta; local error O(h⁵) - Both solvers return a `TODESolution` record with aligned `T` and `Y` arrays ##### Interpolation - `LinearInterp` — piecewise linear with binary-search interval lookup; clamps at endpoints - `LagrangeInterp` — global Lagrange polynomial through all knots - `CubicSplineBuild` / `CubicSplineEval` — natural cubic spline solved via the Thomas tridiagonal algorithm; exact at every knot ##### NumericsLib test coverage 39 new tests in `tests/TestNumericsLib.pas` verify analytically known results: `√2`, the Dottie number, `∫x³ dx = 0.25`, exponential ODE exact solution, harmonic oscillator, and spline exactness at knots. #### EngineeringLib.Signal — FFT and FIR filter design (replacing stubs) ##### FFT (Cooley-Tukey radix-2 DIT) - `FFT(var RealPart, ImagPart; Inverse)` — in-place FFT/IFFT; length must be a power of 2 - `CalculateFFT` — real input → complex spectrum; auto-pads to next power of 2 - `CalculateIFFT` — complex spectrum → real signal - `CalculateFFTMagnitudePhase` — complete N-bin magnitude and phase spectra ##### FIR Filter Design (windowed-sinc) - `DesignFIRLowPass(CutoffFreq, Order, WindowType)` — normalised cutoff in (0, 0.5) - `DesignFIRHighPass(CutoffFreq, Order, WindowType)` — spectral inversion of low-pass - `DesignFIRBandPass(LowCutoff, HighCutoff, Order, WindowType)` — difference of two low-pass filters - `DesignFIRBandStop(LowCutoff, HighCutoff, Order, WindowType)` — notch/band-reject filter - `ApplyFIRFilter(Signal, Coeffs)` — direct-form convolution; output length = N + M − 1 All FIR designs produce symmetric (linear-phase) coefficients. Low-pass is normalised to unit DC gain; high-pass and band-stop use spectral inversion. ##### Signal test coverage `tests/TestEngineeringLib_Signal.pas` expanded from 11 to 52 tests covering FFT linearity, Parseval's theorem, round-trip IFFT accuracy, DC/Nyquist correctness, FIR coefficient symmetry, and `ApplyFIRFilter` impulse response. ### Changed - `EngineeringLib.Signal` — `CalculateFFT` and `CalculateFFTMagnitudePhase` were stubs that raised `Exception` unconditionally; both are now fully implemented. - `tests/TestRunner.lpr` — `TestNumericsLib` registered alongside the existing test suites. --- ## [1.0.1] - 2026-04-16 ### Fixed - `AlgebraLib.Matrices` — `TMatrixKit.IsPositiveDefinite` previously used an insufficient check (determinant > 0 and positive diagonal elements). It now attempts a Cholesky factorisation; success is the definitive test for symmetric positive definite matrices. - `AlgebraLib.Matrices` — `TMatrixKit.IsPositiveSemidefinite` had the same flaw. It now computes all eigenvalues via `EigenDecomposition` and checks that none are less than −1e-9. - `AlgebraLib.Matrices` — `TMatrixKit.Cholesky` previously called `IsPositiveDefinite` as a pre-check, creating a circular dependency after the above fix. The guard is now an inline check: if the diagonal term under the square root is negative, `EMatrixError` is raised immediately. ### Performance - `AlgebraLib.Matrices` — `TMatrixKit.Determinant` replaced recursive cofactor expansion (O(n!)) with LU-based calculation (O(n³)). - `AlgebraLib.Matrices` — `BLOCK_SIZE` increased from 4 to 64 for better L1 cache utilisation. - `AlgebraLib.Matrices` — `TMatrixKit.Multiply` now spawns parallel worker threads for matrices with 64 or more rows. --- ## [1.0.0] - 2026-04-14 ### Changed - The math modules (`TStatsKit`, `TFinanceKit`, `TMatrixKit`, `TTrigKit`, and supporting units) have been separated from [tidykit-fp](https://github.com/ikelaiah/tidykit-fp) into this standalone monorepo. - Source reorganised into focused sub-libraries: `MathBase`, `AlgebraLib`, `FinanceLib`, `EngineeringLib`, and `StatsLib`, each with its own `README.md`. - Unit namespaces updated to match the new library structure (e.g. `MathBase.SharedTypes`, `AlgebraLib.Matrices`, `StatsLib.Stats`). - `EngineeringLib` expanded with `EngineeringLib.FluidDynamics`, `EngineeringLib.Thermodynamics`, `EngineeringLib.Signal`, and `EngineeringLib.UnitConversion`. - `MathBase` expanded with `MathBase.Trigonometry`. ### Removed - All non-math modules (Strings, FS, DateTime, JSON, Logger, Request, Crypto, Archive) are no longer part of this repository; they remain in tidykit-fp. --- ## [0.1.5] - 2025-04-21 (tidykit-fp era) ### Added - Ubuntu 24.04.02 compatibility for DateTime and FS modules. - Automatic test environment detection for the HTTP request module. - HTTP fallback mechanism for testing HTTPS endpoints when OpenSSL is unavailable. - Cross-platform SSL/TLS initialisation support for HTTP requests. ### Fixed - File timestamp handling issues on Unix systems. - Path normalisation for cross-platform compatibility. - OpenSSL initialisation and error handling on Linux systems. --- ## [0.1.0] - 2025-03-13 (tidykit-fp era — initial release) ### Added - `TStatsKit` — statistical calculations. - `TFinanceKit` — financial mathematics. - `TMatrixKit` — matrix operations with decompositions. - `TTrigKit` — trigonometric functions. - JSON operations, logging, cryptography (SHA3, SHA2, AES-256), archive, HTTP client.