mathlib-fp 1.2.3 release notes

Release date: 2026-07-21

Release status: Patch release

mathlib-fp 1.2.3 improves the correctness and numerical robustness of existing operations. It does not introduce a new mathematical domain or require a wrapper, DLL, binary SDK, or third-party runtime library.

Highlights

• Corrected the shared Student-t CDF formula to use the required df/2 incomplete-beta shape parameter. This also corrects statistical p-values calculated through that helper.
• Reworked the scalar special-function kernel with a higher-accuracy Lanczos log-gamma, cancellation-resistant beta evaluation, convergence-checked incomplete-beta continued fractions, and incomplete-gamma normal tails.
• Normal, lognormal, beta, Student-t, and F survival functions now calculate representable upper-tail probabilities directly instead of subtracting a rounded CDF from one.
• F-distribution argument construction avoids overflowing df1*x when the final probability remains representable.
• Hyperbolic and inverse-hyperbolic functions preserve small arguments and avoid avoidable cancellation or intermediate overflow.
• Two-dimensional vector magnitude now uses a scaled hypotenuse calculation, keeping results such as the magnitude of (1e308, 1e308) finite.
• Invalid low-level special-function shape parameters return NaN predictably, and iterative kernels do not silently return unconverged partial results.
• Normal-tail approximations used by statistical tests follow the same direct- tail path as the probability-distribution APIs.
• Kolmogorov-Smirnov empirical CDF steps now use explicit double-precision fractions, avoiding compiler-dependent single-precision evaluation.

Numerical behaviour changes

Applications may observe more accurate results, especially for:

• Student-t CDFs and p-values;
• normal probabilities several standard deviations into a tail;
• beta, Student-t, and F survival probabilities close to zero;
• beta functions with large or asymmetric parameters;
• tiny hyperbolic arguments and values close to inverse-hyperbolic boundaries;
• vector magnitudes containing very large or very small components.

Code that depended on the previous inaccurate Student-t formula, rounded-tail subtraction, or unconverged special-function values should update its expected reference values.

Compatibility

• No public Pascal unit, type, method, property, exception, or parameter was added, renamed, or removed.
• Existing source code compatible with 1.2.2 remains source-compatible with 1.2.3.
• Correctness fixes intentionally change affected numerical results.
• Requirements remain Free Pascal 3.2.2 or later and Lazarus 4.8 or later when using the optional Lazarus package.
• The implementation remains native Object Pascal with no required external runtime dependency.

No application migration is required unless tests or stored results encode the previous numerical approximations.

Quality assurance

• Normal, optimized, runtime-checked, and heap-traced Win64 builds pass all 798 tests with zero errors and zero failures; the heap-traced run reports zero unfreed memory blocks.
• The complete 798-test suite also passes with the native Win32 compiler.
• All 14 runnable examples compile and run.
• The representative benchmark runner compiles and completes.
• The mathlib_fp Lazarus package compiles for Win64 and Win32.
• The Lazarus package metadata reports version 1.2.3.
• New tests cover published reference values, symmetry and complement identities, invalid inputs, convergence outcomes, tiny arguments, very large vector components, representable extreme distribution tails, and portable Double evaluation across FPC targets.
• All local Markdown targets resolve, and git diff --check reports no whitespace errors.

For the complete change list, see the 1.2.3 changelog. The broader native Free Pascal numerical-computing direction is described in the project roadmap.

Bug reports and contributions are welcome through the GitHub repository. For security reports, follow the security policy.

mathlib-fp is distributed under the MIT License.