Release 1.0.0 (#45)

* bump version

* add weights argument (#29)

* remove unused variable (#26)

* fix edf computation for deg = 2

* bump version + update docs

* remove package cache on appveyor

* weighted bandwidth selection

* weighted fitting

* adjust R interface

* add unit tests

* a few more sanity checks

* update docs

* update docs

* fix devtools install link

* stabilize weighted quantiles

* ensure scale != 0 in case of too many ties

* remove stripping of debug symbols

* prepare release

* update docs

* allow already jittered input to be treated as continuous

* some improvements to interpolation (#32)

* use sample quantiles as grid points

* fix unweighted quantile computation

* improve safety of integral computations.

* make sure that weighted version is consistent

* improve grid extensions

* fix boundary handling for bounded domain

* fix interpolation bug when grid distance is small

* install gsl

* extend grid a bit further

* fix bug for  weighted quantile with zero weight

* decrease lower bound for interpolation, related to vinecopulib/rvinecopulib#157

* More numerical stability (#35)

* hybrid quantile/equidist grid

* normalize weights

* fix weighted influence and use it as safeguard

* avoid double computation of influence fot deg = 0

* interpolant positivity and smooth extrapolation (#37)

* switch to linear interpolation if result is negative

* use positive constraints for spline coefficients

* Gaussian tail for extrapolation

* optimal bandwidths for all degrees (#38)

* missing inline

* fix levels in pkde1d

* stabilize edf computation

* tabs = 2 spaces (#39)

* simplified, deterministic jittering (#40)

* tabs = 2 spaces

* simplified, deterministic jittering

* prepare standalone c++ version with interface (#41)

* rename lpdens -> kde1d

* use namespaces

* bw selection in Kde1d class

* jittering + nan handling in cpp

* statistical functions

* constructor from interpolation grid

* all discrete functionality in cpp

* kde1d dir structure

* seperate wrappers from interface

* adapt R code

* go to office

* fix nans in grid

* fix boundary transform order

* refactor unit tests

* Fast integration (#42)

* rename lpdens -> kde1d

* use namespaces

* bw selection in Kde1d class

* jittering + nan handling in cpp

* statistical functions

* constructor from interpolation grid

* all discrete functionality in cpp

* kde1d dir structure

* seperate wrappers from interface

* adapt R code

* go to office

* fix nans in grid

* fix boundary transform order

* refactor unit tests

* unnormalized algorithm

* faster integration by sorting input

* Fit models with the fast fourier transform (#43)

* isolate KdeFFT class

* almost done with fft

* proper weights in edf calculation

* fix grid related issues

* eliminate dead code

* safer integration

* reorder arguments for the feels

* prepare release (#44)

* Release 0.4.0 (#31)

* bump version

* add weights argument (#29)

* remove unused variable (#26)

* fix edf computation for deg = 2

* bump version + update docs

* remove package cache on appveyor

* weighted bandwidth selection

* weighted fitting

* adjust R interface

* add unit tests

* a few more sanity checks

* update docs

* update docs

* fix devtools install link

* stabilize weighted quantiles

* ensure scale != 0 in case of too many ties

* remove stripping of debug symbols

* prepare release

* update docs

* bump version

* update NEWS

* update docs

* update NEWS and DESCRIPTION

* update API docs

* backwards compatibility with bad code in rvineocpulib *sigh*

* initialization order warning

* typo in API doc

* CRAN comments

* API docs one more time

.gitignore

@@ -4,3 +4,4 @@
 .Ruserdata
 revdep
 /revdep/.cache.rds
+.vscode

.travis.yml

@@ -27,6 +27,7 @@ before_install:
   - sudo apt-get install gcc-5 g++-5 gfortran-5
   - sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-5 100
   - sudo update-alternatives --install /usr/bin/gfortran gfortran /usr/bin/gfortran-5 100
+  - sudo apt-get install libgsl-dev
 
 repos:
   CRAN: http://cran.rstudio.com

DESCRIPTION

@@ -1,7 +1,7 @@
 Package: kde1d
 Type: Package
 Title: Univariate Kernel Density Estimation
-Version: 0.4.0
+Version: 1.0.0
 Authors@R: c(
         person("Thomas", "Nagler",, "mail@tnagler.com", role = c("aut", "cre")),
         person("Thibault", "Vatter",, "thibault.vatter@gmail.com", role = c("aut"))
@@ -20,13 +20,12 @@ LinkingTo:
     Rcpp,
     RcppEigen
 Imports:
-    cctools, 
     graphics,
     Rcpp, 
     qrng,
     stats,
     utils
-RoxygenNote: 6.1.1
+RoxygenNote: 7.0.0
 Suggests: 
     testthat
 URL: https://github.com/tnagler/kde1d

NAMESPACE

@@ -6,13 +6,12 @@ S3method(plot,kde1d)
 S3method(print,kde1d)
 S3method(summary,kde1d)
 export(dkde1d)
+export(equi_jitter)
 export(kde1d)
 export(pkde1d)
 export(qkde1d)
 export(rkde1d)
 importFrom(Rcpp,sourceCpp)
-importFrom(cctools,cont_conv)
-importFrom(cctools,expand_as_numeric)
 importFrom(graphics,lines)
 importFrom(graphics,plot)
 importFrom(qrng,ghalton)

NEWS.md

@@ -1,3 +1,28 @@
+# kde1d 1.0.0
+
+NEW FEATURES
+
+  * optimal plug-in bandwidth selection for all polynomial degrees (#38).
+
+  * avoid randomness through simplified, deterministic jittering, see 
+  `equi_jitter()` (#40).
+
+  * removed dependency `cctools`.
+
+  * headers in `inst/include` can be used as standalone C++ library with 
+    convenience wrappers for R (#41).
+
+  * (several times) faster `pkde1d()`, `qkde1d()`, and `rkde1d()` due to
+    a more clever algorithm for numerical integration (#42).
+
+  * faster `kde1d()` thanks to the Fast Fourier Transform (#43).
+
+BUG FIXES
+
+  * improvements to numerical stability, inter- and extrapolation (#32, #35, 
+  #37).
+
+
 # kde1d 0.4.0
 
 NEW FEATURE

R/RcppExports.R

@@ -3,7 +3,9 @@
 
 #' fits a kernel density estimate and calculates the effective degrees of
 #' freedom.
-#' @param x vector of observations.
+#' @param x vector of observations; catergorical data must be converted to
+#'   non-negative integers.
+#' @param nlevels the number of factor levels; 0 for continuous data.
 #' @param bw the bandwidth parameter.
 #' @param xmin lower bound for the support of the density, `NaN` means no
 #'   boundary.
@@ -13,49 +15,35 @@
 #' @return `An Rcpp::List` containing the fitted density values on a grid and
 #'   additional information.
 #' @noRd
-fit_kde1d_cpp <- function(x, bw, xmin, xmax, deg, weights) {
-    .Call('_kde1d_fit_kde1d_cpp', PACKAGE = 'kde1d', x, bw, xmin, xmax, deg, weights)
+fit_kde1d_cpp <- function(x, nlevels, bw, mult, xmin, xmax, deg, weights) {
+    .Call('_kde1d_fit_kde1d_cpp', PACKAGE = 'kde1d', x, nlevels, bw, mult, xmin, xmax, deg, weights)
 }
 
 #' computes the pdf of a kernel density estimate by interpolation.
 #' @param x vector of evaluation points.
-#' @param R_object the fitted object passed from R.
+#' @param kde1d_r the fitted object passed from R.
 #' @return a vector of pdf values.
 #' @noRd
-dkde1d_cpp <- function(x, R_object) {
-    .Call('_kde1d_dkde1d_cpp', PACKAGE = 'kde1d', x, R_object)
+dkde1d_cpp <- function(x, kde1d_r) {
+    .Call('_kde1d_dkde1d_cpp', PACKAGE = 'kde1d', x, kde1d_r)
 }
 
 #' computes the cdf of a kernel density estimate by numerical integration.
 #' @param x vector of evaluation points.
-#' @param R_object the fitted object passed from R.
+#' @param kde1d_r the fitted object passed from R.
 #' @return a vector of cdf values.
 #' @noRd
-pkde1d_cpp <- function(x, R_object) {
-    .Call('_kde1d_pkde1d_cpp', PACKAGE = 'kde1d', x, R_object)
+pkde1d_cpp <- function(q, kde1d_r) {
+    .Call('_kde1d_pkde1d_cpp', PACKAGE = 'kde1d', q, kde1d_r)
 }
 
 #' computes the quantile of a kernel density estimate by numerical inversion
 #' (bisection method).
 #' @param x vector of evaluation points.
-#' @param R_object the fitted object passed from R.
+#' @param kde1d_r the fitted object passed from R.
 #' @return a vector of quantiles.
 #' @noRd
-qkde1d_cpp <- function(x, R_object) {
-    .Call('_kde1d_qkde1d_cpp', PACKAGE = 'kde1d', x, R_object)
-}
-
-#' @param x vector of observations
-#' @param grid_size number of equally-spaced points over which binning is
-#' performed to obtain kernel functional approximation
-#' @param weights vector of weights for each observation (can be empty).
-#' @return the selected bandwidth
-#' @noRd
-select_bw_cpp <- function(x, bw, mult, discrete, weights) {
-    .Call('_kde1d_select_bw_cpp', PACKAGE = 'kde1d', x, bw, mult, discrete, weights)
-}
-
-quan <- function(x, a, w) {
-    .Call('_kde1d_quan', PACKAGE = 'kde1d', x, a, w)
+qkde1d_cpp <- function(p, kde1d_r) {
+    .Call('_kde1d_qkde1d_cpp', PACKAGE = 'kde1d', p, kde1d_r)
 }
 

R/jitter.R

@@ -0,0 +1,45 @@
+#' Conditionally equidistant jittering
+#'
+#' Converts ordered variables to numeric and Adds deterministic uniform noise.
+#' See *Details*.
+#'
+#' Jittering makes discrete variables continuous by adding noise. This simple
+#' trick allows to consistently estimate densities with tools designed for the
+#' continuous case (see, Nagler, 2018a/b). The drawback is that estimates are
+#' random and the noise may deteriorate the estimate by chance.
+#'
+#' Here, we add a form of deterministic noise that makes estimators well
+#' behaved. Tied occurences of a factor level are spread out uniformly
+#' (i.e., equidistantly) on the interval \eqn{[-0.5, 0.5]}. This is similar to
+#' adding random noise that is uniformly distributed, conditional on the
+#' observed outcome. Integrating over the outcome, one can check that the
+#' unconditional noise distribution is also uniform on \eqn{[-0.5, 0.5]}.
+#'
+#' Asymptotically, the deterministic jittering variant is equivalent to the
+#' random one.
+#'
+#' @param x observations; the function does nothing if `x` is already numeric.
+#'
+#' @export
+#'
+#' @references
+#' Nagler, T. (2018a). *A generic approach to nonparametric function estimation
+#' with mixed data.* Statistics & Probability Letters, 137:326–330,
+#' [arXiv:1704.07457](https://arxiv.org/abs/1704.07457)
+#'
+#' Nagler, T. (2018b). *Asymptotic analysis of the jittering kernel density
+#' estimator.* Mathematical Methods of Statistics, in press,
+#' [arXiv:1705.05431](https://arxiv.org/abs/1705.05431)
+#'
+#' @examples
+#' x <- as.factor(rbinom(10, 1, 0.5))
+#' equi_jitter(x)
+equi_jitter <- function(x) {
+  if (is.numeric(x))
+    return(x)
+  x <- as.numeric(x)
+  tab <- table(x)
+  noise <- unname(unlist(lapply(tab, function(l) -0.5 + 1:l / (l + 1))))
+  s <- sort(x, index.return = TRUE)
+  (s$x + noise)[rank(x, ties.method = "first", na.last = "keep")]
+}