Switch build to use pyproject.toml, include only package directory. Updates to tests for numpy 2.0. Some housekeeping (since v1.2.2).
Decrease ftol used by the FB5 minimizer. Previously, for some test cases when the initial seed (moment estimate) is close to optimal the SLSQP minimizer did not find an improved solution.
Update rng seed handling to allow the rvs method to take an optional seed. By default, seed=False which preserves the previous caching behavior. Passing a seed will clear the cache and generate new/reproducible rvs.
Fix a bug where OverflowError would not be caught properly in normalize for normalization edge cases near float max.
Patch an edge case in FB8Distribution.contour calculation when beta is 0.
Faster implementation of the FB8 normalization calculation using heapq. The idea is to first run an increasingly coarser grid search over the indices to find the approximate-maximum summand, then start with a 3D cube around that point. Contributions from its six sides and their next-step coordinates are placed in a heap such that the next-largest contribution is summed next.
Optimize series summations by reducing repeated special function calls. Fix a bug and catch an edge.
Implement gradient calculations for mle fitting. This includes series computation of derivatives of the FB8 normalization wrt its parameters.
Minor optimizations by switching to np.mgrid and np.moveaxis.
Run Travis.ci tests for multiple python versions. Updates and additions to example scripts in paper/fig.py
Python 3 support.
Implement series calculation for FB8 normalization and make that the default. Fall back is numerical integration. A closed form approximation for the FB6 normalization constant is also implemented, which allows for computing the log-pdf for large values of kappa and beta.
First working version with FB8 distribution implemented. Its normalization constant is computed using numerical integration.