NEWS

smoothemplik 0.0.17 (2025-10-29)

Defined EL0() as the univariate and EL1() as the multi-variate wrapper, subsumed by EL()
Added Extrapolated EL for applications where the convex hull can be violated
Re-implemented adjusted (2 variants) and balanced EL
Added Bartlett factor computation
Improved brentZero() root search, added two new extensions rules: "left" and "right"
Improved the damped Newton search, added several stopping criteria to prevent stalling

Fixed a bug in the Taylor expansion for the case of convex hull condition failures
Removed unfinished functions for optimisation and constrained optimisation
Implemented adaptive bandwidths at each point of the evaluation grid for better smoothing
Added a draft version of the vignette showcasing the use of adaptive kernels for smoothing
Implemented a Wald-like Taylor expansion to allow mu to be outside the convex hull in weightedEL()
Sped up the fourth-order triangular kernel

Implemented Taylor expansion to allow mu to be outside the convex hull in weightedEL(), as in Owen (2013)
Replaced uniroot() with a C++ version of Brent's zero search for speed
Improved handling of influential observations in LSCV

Fixed a bug in prepareKernel() where a valid y vector with attributes would not pass the check.
Implemented a more accurate check for lambda being close to the boundary based on the relative search interval length in weightedEL().
weightedEL() preserves the names of the input vector in wts.
Sped up ctracelr() by using the previous lambda value in the search (~4 times faster).
The output of mllog() now has column names because it was confusing without them.
The output of svdlm() is now a vector, not a 1-column matrix.
Replaced certain instances of sapply() with vapply() in smoothing functions.
Added unit tests for some functions.

Feature: support for weighted kernel density and regression observation.
Feature: support for sparse weight matrices via sparseKernelWeightsCPP() to save memory.
Feature: observation de-duplication to speed up the non-parametric functions.
Feature: low-level C++ parallelisation and chunking to limit maximum RAM use (via RcppParallel).
Feature: leave-one-out kernel smoothing support for custom output grids.
Reworked mixed kernel density and regression workflow making use of the block data structure.
Bug fix: now there are time savings in multi-core mixed estimation.
Bug fix: in DCV, duplicated observations were not merged (now the LOO is the true LOO: duplicates of the same observation are also left out).
Improved the initial value choice for cross-validation (using a log-scale grid around the rule-of-thumb value).
Sped up C++ kernel functions with RcppArmadillo (20--50% speed gains through better iterations, code structure, and condition checks).

Added a vignette on non-parametric methods.
Added 4th-order C++ versions of all kernels (for bias reduction) and their convolutions.
Auto-detect cross-validation type (density or least-squares) based on the input.
Changed the default behaviour of Silverman's rule of thumb: use a robust estimator of SD (IQR/1.34).
Prepared a stub for discontinuous densities.
Moving the gradient-related functions to a new package, pnd.

Rewrote the C++ smoothers using RcppArmadillo for speed-up, refactored the kernel-related code.
Feature: Support for the 2nd-order uniform, triangular, Epanechnikov, and quartic kernel.

Feature: added functions for parallelised numerical differentiation.
Rewrote multi-variate weighted empirical likelihood functions to allow for Taylor approximations of the empirical likelihood function of any order.

Feature: getSELWeights now renormalises the weights to unity after trimming (by default, can be overridden via renormalise = FALSE).