mathlib-fp 1.2.0 release notes

Release date: 2026-07-18 Release status: First public release

mathlib-fp 1.2.0 is the first public release of the standalone mathlib-fp project. It provides native Free Pascal units for scientific, statistical, financial, engineering, numerical, optimization, time-series, machine-learning, and geometry work without third-party runtime dependencies.

What is included

The release contains 12 focused domains:

See the documentation index for the API guides and the examples for runnable programs.

Highlights

• Real symmetric eigendecomposition and supported real 2×2 nonsymmetric cases use residual-based convergence checks; unsupported nonsymmetric or complex spectra raise EMatrixError.
• Fractional matrix powers use symmetric eigendecomposition and require a symmetric positive-definite matrix; integer powers use exponentiation by squaring.
• Matrix multiplication uses a bounded worker count, propagates worker errors, and falls back safely when a Unix program has no thread manager installed.
• Seeded random and bootstrap overloads provide reproducible results without changing global random state.
• Machine-learning entry points consistently reject empty, ragged, non-finite, mismatched, and out-of-range inputs with EMLError.
• Financial APIs apply documented decimal rounding consistently, reject undefined ratios, and calculate positive or negative IRRs with bracketed bisection instead of returning a fixed initial guess.
• FinanceLib.Bonds and FinanceLib.NPV provide focused import paths with directly nameable bond schedule and NPV cash-flow aliases; their formulas remain centralized in FinanceLib.Interest.
• Engineering APIs have shared, domain-specific exception types through EngineeringLib.Common; the focused Velocity and Pressure units also expose directly nameable aliases for their calculation class and exception type.
• Signal processing includes complete-spectrum FFT/IFFT, symmetric windowed- sinc FIR designs for even or automatically adjusted odd orders, and the shared MathBase numeric-array type.
• Fluid and thermodynamic APIs validate solver, pressure, temperature, and specific-heat-ratio domains with deterministic Engineering exceptions.
• Pump head uses the Bernoulli relation with explicit inlet and outlet velocities, avoiding an ambiguous pre-public API.
• Unit conversion no longer silently treats unknown units as lengths: non-Try APIs raise EUnitConversionError, while Try... APIs return False. Its formatting, parsing, enumeration, base-unit, and shortcut APIs are fully documented, including exact symbols and fixed month/year durations.
• Matrix inverse, LU, rank, and exponential paths have scale-aware correctness fixes backed by reference, reconstruction, and algebraic property tests.
• Statistics sorting and convex-hull point ordering use O(n log n) algorithms; a representative benchmark runner covers both paths and dense multiplication.
• Statistical inference now returns exact small untied Mann-Whitney p-values, calibrated approximate K-S and Shapiro-Wilk p-values, and pooled-variance Cohen's d, with published/reference-value regression tests.
• Time-series frequency bins and ARIMA MA/integration forecasts have corrected mappings and reconstruction, including non-power-of-two and differenced polynomial reference cases.
• Linear regression uses centered Householder QR, while PCA reports iteration counts, enforces convergence, re-orthogonalises components, and handles rank-deficient covariance deterministically.
• Numerical root finders offer detailed convergence records; scalar root and optimization solvers raise on iteration exhaustion instead of silently returning an unconverged estimate. Linear programming reports a termination status in addition to its compatibility Feasible flag.
• Geometry, combinatorics, trigonometry, interpolation, distributions, and optimizer validation received focused correctness and domain regression tests.
• Win32-specific release validation corrected sieve index portability, significant-digit tie rounding, LU triangular cleanup, and deferred floating-point overflow reporting.

For the complete list of additions and fixes, see the changelog.

Requirements and installation

• Free Pascal 3.2.2 or later.
• Lazarus 4.8 or later when using the optional Lazarus package.
• No third-party runtime dependencies.

Download a source archive from the GitHub release, or clone the repository. Add src/ to the compiler unit search path and keep compiler output in a separate directory:

mkdir -p lib

fpc -Fusrc -FUlib my_program.pas

Lazarus users can install packages/lazarus/mathlib_fp.lpk, or add src/ under Project Options → Compiler Options → Paths → Other Unit Files.

Compatibility and migration notes

• This is the first supported public release. Versions 1.1.x and older were pre-public development versions and are not supported.
• The Lazarus package and registration unit are named mathlib_fp. Remove any pre-release package installation from Lazarus before installing this release.
• The user-facing project and release name remains mathlib-fp; the underscore is used only where Pascal identifiers cannot contain a hyphen.
• On Unix, include cthreads before other units in applications that require threaded matrix multiplication. The implementation falls back to non-threaded execution when a thread manager is unavailable.

Validation

The release candidate was validated with Free Pascal 3.2.2 on Linux and Windows:

• 788 automated tests pass with zero failures on both Win64 and Win32.
• Normal, optimized, checked, and heap-traced test builds pass.
• All 12 example programs compile and run.
• The README quick-start builds and runs successfully.
• Windows CI builds the mathlib_fp package and runs all 788 tests with native Win64 and Win32 Lazarus 4.8 toolchains.
• The package and all 788 tests also pass in a manual Windows 11 installation of Lazarus 4.8 with its bundled Free Pascal compiler.
• The benchmark runner compiles and completes representative workloads.
• A clean source-archive installation repeats the tests, examples, quick-start, benchmark compilation, and Lazarus-package build.
• Clean normal and UTF-8 builds complete without compiler warnings.

Support, security, and license

Bug reports and contributions are welcome through the GitHub repository. Do not disclose vulnerability details in a public issue; follow the security policy to submit a private report.

mathlib-fp is distributed under the MIT License.