Changelog#

1.1.0 (2026-06-05)#

Added SciPy-compatible support for jac=True in minimize_lbfgsb and prepare_scalar_function. When jac=True, the objective function may now return both the objective value and gradient as fun(x) -> (f, g).
Added memoization support for objective functions returning (f, g) to avoid duplicate evaluations when both the function value and gradient are requested at the same point.

Improved typing annotations for objective functions, gradients, finite-difference options, and scalar-function wrappers.
Improved compatibility with static type checkers, including ty.
Improved Python 3.7 compatibility for typing-related constructs.
Updated line-search handling to be more defensive against non-finite trial points, non-finite directional derivatives, and invalid step lengths.
Improved numerical robustness of line-search utilities, including safer dot products, safer squared-norm computations, and safer maximum step-length detection.
Replaced mutable default work arrays in the line-search routine with None defaults initialized inside the function.

Fixed scalar objective normalization in ScalarFunction so that objective values are consistently converted to plain Python float values.
Fixed support for complex-step finite differences by preserving complex objective values during numerical differentiation.
Fixed typing issues caused by broad NumPy scalar/object unions in objective function evaluation.
Fixed potential use of uninitialized line-search step variables.

This is the first stable release with a non negligeable number of new features and performance improvement:

Checkpointing (start and stop mechanisms)
On-the-fly modification of the objective function
Added benchmarks with scipy
Provide an optional numba jit implementation for the expensive parts of the code.
Enhanced documentation and tests.