Release Notes

Unreleased

  • A new weights keyword argument has been added to the kde function to support weighted data sets. As a consequence, the API of multiple interfaces have been changed:
    • bandwidth has gained a weights keyword.
    • The init method returns three values (data, weights, and info) instead of just two.
    • The estimate function takes a mandatory weights positional argument.

v0.6.0 — 2024 Dec 31

  • Public functions have been declared using Julia v1.11+'s public keyword.
  • The value of the bin-center for the zero-width singleton histogram has been fixed.
  • Implement more accurate histogram (and linear) binning calculations. For the HistogramBinning case, it is now possible to precisely bin range(lo, hi, nbins) values into their corresponding bins (whereas previously values may be counted incorrectly one bin too low due to rounding in the floating point calculations).
  • The implementation has been modified to support unitful quantities (without adding a new package dependency) via careful consideration and application of appropriate factors of one and/or oneunit (and relaxing type constraints or adding new type parameters to structs, where necessary). Given a vector with units u, the density object's fields (K.x, K.f) have units (u, u^-1), respectively, in correspondence with interpreting any integrated range to be a unitless probability (i.e. probability = sum(K.f[a .< K.x .< b]) * step(K.x)).
  • The package no longer depends on Roots.jl (used by the ISJ bandwidth estimator); instead, an implementation of Brent's Method is now included here directly. This not only decreases the dependence on external packages but also reduces both package load time and the precompiled package image size.
  • The documentation has generally been improved:
    • A new "Showcase" section has been added to showcase examples of density estimation with this package.
    • A User Guide has been started to give a brief introduction to installing and using the package.
    • The documentation now includes release notes.

v0.5.0 — 2024 Nov 21

  • Have the ISJ bandwidth estimator fallback to Silverman rule automatically when it fails to converge. This is expected to happen for very flat (closed) distributions, so it's not as rare of an occurrence as originally understood.
  • Add package extension to aid in plotting with Makie.jl.
  • Begin adding more extensive documentation to the package:
    • Add a README to give brief justification for the package.
    • Describe the package extensions available and what features they provide.
    • Demonstrate the impact of the different stages of the estimator pipeline by comparing each stage for several example distributions.

v0.4.0 — 2024 Sep 09

  • Add an interface method boundary which can be overloaded to implement mechanisms for automatically determining appropriate boundary conditions.
    • The two built-in methods are to convert from symbols to enum (e.g. :open to Open) and to infer the boundary conditions from a 2-tuple of finite/infinite real values.
  • Add an interface method bounds (and eponymous keyword argument to kde) which can be overloaded to implement mechanisms for automatically both the boundary condition (keyword boundary) and limits (keywords lo and hi) from an arbitrary value.
  • Store more information within the UnivariateKDEInfo structure. The init uses the new fields to pass relevant parameters to later stages of the estimator pipeline.
  • On Julia v1.9+, new extension packages have been added to integrate with external packages:
    • The aforementioned boundary and bounds methods have been specialized for univariate distributions from Distributions.jl
    • The UnicodePlots.jl package, if loaded, is used to visualize the kernel density estimate at the terminal.
  • Increase the automatic bandwidth determined by the chosen bandwidth estimator for higher-order estimators (such as MultiplicativeBiasKDE) which have a lower level of bias. (See Lewis [2], §E, Eqn 35 and Footnote 10 for further details.)
  • Rename the boundary condition enum from Cover to Boundary.
  • Fix syntax or usage errors that broke compatibility with Julia v1.6.

A. Lewis. GetDist: a Python package for analysing Monte Carlo samples (2019), arXiv:1910.13970.

v0.3.0 — 2024 Jul 21

This release adds the ISJ bandwidth estimator described in Botev et al. [4] and uses it by default, as it is more capable of dealing both with non-Gaussian distributions and respects the complexity/nature of bounded domains.

  • Add an implementation of the Improved Sheather-Jones bandwidth estimator.
  • Rename the interface method for bandwidth estimators to bandwidth.
  • Consolidate data and option pre-processing into the init method, which is the first step in the density estimation pipeline.

Z. Botev, J. Grotowski and D. Kroese. Kernel density estimation via diffusion. The Annals of Statistics 38 (2010), arXiv:1011.2602.

v0.2.0 — 2024 Jul 19

  • Require Julia v1.6+.
  • Improved docstrings throughout.
  • Added framework for building documentation with Documenter.jl.
  • Migrate specific density estimation implementations to be methods of the (new) estimate interface function (rather than overloading kde).
  • Fix handing of edge-case where a constant vector is given. For closed boundary conditions, the result is a zero-width singleton bin.
  • Fix error in widening of the KDE range based on the kernel bandwidth.

v0.1.0 — 2024 Jul 13

Initial release supports: