Change log file for ks
1.9.2
-Changed binning=FALSE to binned=binned for Hpi(,pilot="dscalar"),
Hpi.diag(,pilot="dscalar").
-Fixed bug in binning behaviour in gdscalar().
1.9.1
-Fixed typos in help files
-Added new classes "kcopula" and "kcopula.de" for output from kcopula and
kcopula.de to distinguish them from "kcde" and "kde" objects.
-Exported matrix.sqrt().
-Added "exp" option for make.grid.ks().
1.9.0
-Added efficient recursive versions for dmvnorm.deriv(), Sdr(), Sdrv(),
nur(), nurs(), mur(), Qr() from Chacon & Duong (2014) Statist Comput.
-Fixed bug in Hscv(,binned=TRUE), Hscv.diag(,binned=TRUE) which was still
computing unbinned estimators.
-Fixed bug in 1-d KDA plot.
-Added sensitivity, specifity as output to compare().
-Made small changes to default selectors to be more consistent across
selectors.
-Fixed bug in point colour in rug plot for plotkda.1d().
-Added Hpi.diag.kcde().
-Added Lpdiff() (Lp distance for two functions) and copula.grid (true copula
evaluated on a grid).
-Fixed small bug in plotmixt(,draw=FALSE) to actually not draw plots.
-Added predict method for kde objects to replace kde.approx().
-Added option to compute 1-d KDE supported on [0,1] kde(,unit.interval=TRUE)
which calls kde.unit.interval().
-Changed default axes limits when plotkde.3d(, drawpoints=FALSE) from data
range to mean of KDE evaluation range.
-Fixed bug in default pilot selector for d>3 data in kda().
-Changed ad hoc argument matching to match.arg().
-Fixed bug in last line of lscv.mat().
-Added binned estimation to Hbcv(), Hbcv.diag().
-Added default binning flag function default.flag().
1.8.13
-Added boundary density estimator kde.boundary() for compactly supported
data.
-Added kernel density of copula nd copula density, i.e. kcopula() and
kcopula.de().
-Fixed small bug in plot.kcde(disp="slice", abs.cont=!missing), and
Hpi.kcde().
-Changed smoothing spline in kroc() to be evaluated on equally spaced
grid.
-Added thinning option for persp plots plot.kde(thin=), plot.kcde(thin=).
1.8.12
-Added kernel estimators for CDF kcde() and ROC curves kroc().
-Added default plug-in bandwidths to kda(), kcde(), kde(), kdde(),
kde.local.test(), kroc(), kde.test().
-Added warning when using non-diagonal bandwidths for binned estimation.
-Added plot and contourLevel methods for kdde objects.
-Modified plotmixt() to include derivatives.
-Added 1-d plug-in selectors hpi(,deriv.order>0).
-Merged kda() and kda.kde() into single kda() function.
-Changed "kda.kde" object class name to "kda".
1.8.11
-Added progress bars to compare.kda.cv(), compare.kda.diag.cv().
-Corrected critical df from d to 1 in kde.local.test().
1.8.10
-Fixed small bug in call to contourLevels(approx=) inside kde().
1.8.9
-Added kde.local.test() for local 2-sample test.
-Replaced foreign call to .C("massdist", package="stats") requested by
B. Ripley by call to .C("massdist1d", package="ks").
-Changed rug plot in plot.kda.kde() to rug-like plot, similar that in
plot.kde.loctest().
1.8.8
-Changed function header of Hpi.kfe() to be more consistent with Hpi().
-Added option Hpi.kfe(, pilot="dscalar") to ensure scale invariance in
p-values. This becomes the default over the previous pilot="unconstr".
-Added 1-d option in kde.test() and its required bandwidth hpi.kfe().
-Modified binned=TRUE option in kde.test() so that it is applied only to
bandwidth selection, and not the test statistic and its p-value.
-Removed default truncation in Hlscv(), Hlscv.diag() for deriv.order=0.
1.8.7
-Further improved speed of kfe(,Sdr.flag=FALSE) by computing unique
partial derivatives.
-Removed unused function dkde.weights() to compute optimal deconvolution
weights, and hence dependence on the kernlab library.
-Changed output from kfe(binned=TRUE) to be vector not 1-row matrix.
1.8.6
-Implemented calcualtion of Lebesgue measure of level sets of contours,
contourSizes().
-Implemented probability contour plot for 1-d KDE plot, i.e. analogue
to existing 2-d, 3-d contour plot(,disp="slice").
-Added recursive computations kfe(,Sdr.flag=FALSE) which don't compute
symmetriser matices explicitly. These are then called in
Hpi(,Sdr.flag=TRUE) and Hscv(,Sdr.flag=FALSE).
-Changed pilot="dunconstr" to direct computation rather than indirect
eta form. This means that Hpi(,pilot="dunconstr", deriv.order=0) and
Hpi(,pilot="unconstr") now give the same result.
-Remove pilot="dsamse" option as this was more computation than
pilot="dscalar" but with little difference in the result.
1.8.5
-Fully unconstrained pilot selectors pilot="dunconstr" for Hscv(),
Hpi() for density derivative estimation.
-Unconstrained Hlscv() selector for density derivative estimation.
1.8.4
-Reinstated psi.ns code (more efficient than eta.kfe.y) and SAMSE
pilot estimators Hpi(, pilot="samse").
-Edited help manual.
1.8.3
-Added computationally efficient density derivative b/w selectors
Hpi(deriv.order=), Hscv(deriv.order=), and their diagonal counterparts
Hpi.diag(), Hlscv.diag().
-Added computationally efficient kernel functional estimators in
eta.kfe.y() used in kde.test().
-New pilot selectors for density derivatives.
-Added abs.cont capability to plot(, disp="filled.contour").
-Removed explicit expressions in psins() for d>2, replaced by
eta.kfe() evaluations.
-Removed psins() and Theta6() evaluations in gsamse and gamse.scv.
-Removed kfold arguments.
1.8.2
-Fixed bug in kde.points.sum() to avoid allocating large matrices for
unbalanced sample sizes for x and eval.points.
-Fixed bug in dmvnorm.deriv.sum() which had excluded last partition
class for double.loop=FALSE.
-Added binned options to kde.test().
-Fixed bug for exact estimation in kfe().
-Added plotting colours as function of z-value in plot.kde(,
disp="persp").
-Added decoupled calculcation for Hlscv().
-Added optim.fun option to select optimiser function in Hpi, Hpi.diag,
Hlscv, Hlscv.diag, Hscv, Hscv.diag().
1.8.1
-Modified p-value calculation for large -ve Z-statistics.
-Fixed bug for binned estimation for unconstrained bandwidths for kde().
1.8.0
-Added density derivative selectors Hpi(,deriv.order=r),
Hlscv(,deriv.order=r) for r>0 from J.E. Chacon.
-Changed vech(H) terms to vec(H) in AMISE estimators.
-Changed default binning gridsize for 3-d data from rep(51,3) to
rep(31,3).
-Added verbose option to b/w selectors (in double sum) for tracking
progress.
-Changed LSCV, SCV selectors optimisation from Nelder-Mead to BFGS.
-Changed Fortran linear bining code to C (and fixed bugs in Fortran
code).
-Added modification to linear binning for boundary points.
-Removed explicit derivatives in BCV selector optimisation.
1.7.4
-Fixed small bug in partitioning method for kde.points.sum().
1.7.3
-Changed partitioning method for dmvnorm.deriv.sum() and
kde.points.sum().
1.7.2
-Changed p-value calculation for kde.test().
1.7.1
-Reinstated single partial derivative of mv normal for scalar variance
matrix dmvnorm.deriv.scalar.sum() for use in AMSE pilot plug-in
selectors.
-More efficient form of kdde().
1.7.0
-Added KDE-based 2-sample test kde.test().
-Modified output of plotmixt().
-Added "double.loop" option to kfe() for large samples - increases
running time, reduces memory.
-Modified dmvnorm.deriv.sum() gto improve memory memory management for
large samples.
-Cleaned up code for plug-in bandwidth selectors and kernel functional
estimators.
-Cleaned up help files.
-Disabled kfold b/w selectors.
1.6.13
-Added flag to automatically compute probability contour levels in kde().
1.6.11
-Added own version of filled contours as option disp="filled.contour2"
and different colours for disp="slice" contours.
1.6.10
-Added k-fold b/w selectors.
1.6.9
-Added approximate option in contourLevels().
-Added kdde() kernel density derivative estimators.
1.6.8
-Added 1-d LSCV selector hlscv().
1.6.7
-Corrected ISE for normal mixtures, from J.E. Chacon.
1.6.6
-Added MISE, AMISE, ISE functions for normal mixtures derivatives.
-Changed internal double sum calculations from J.E. Chacon.
1.6.x
-1-d binned KDE fix from M.P. Wand.
-Streamlined code sharing with feature package (all binning code now
contained only in ks).
-Reorganised and renamed internal bandwidth selection functions,
mstly double sums of normal densities .
1.5.11
-Fixed small bugs in drvkde, vech, Hpi(, pilot="unconstr")
1.5.10
-Added drvkde (kernel density derivative estimator 1-d) from feature
using M.P. Wand's code.
1.5.x
-Added normal mixture (A)MISE-optimal selectors: hamise.mixt,
hmise.mixt, Hamise.mixt, Hmise.mxt.
-Added distribution functions for 1-d KDEs: dkde, pdke, qkde, rkde.
-Added plug-in selectors for 1-d data (exactly the code for dpik from
KernSmooth). For KDE, this is hpi, for KDA, this is hkda(,
bw="plugin").
-Made changes to specifying line colour (col rather than lcol) in
plot.kde, plot.kda.kde and partition class colour (partcol) in
plot.kda.kde.
-Added plot3d() capabilities from rgl to 3-d plot - removing own axes
drawing functions.
-New functions to compute pilot functinal estimators
hat{psi}_r(g). These are exact, and are more efficient than binned
estimators for small samples (~100), and are available in d > 4.
1.4.x
-Vignette illustrating 2-d KDE added
-Binned estimation implemented for KDE with diagonal selectors and
pilot functional estimation with diagonal selectors.
-Filled contour plots added as disp=filled option in plot.kde().
-compare.kda.cv() and compare.cv() modified to improve speed.
-Hscv.diag() and Hbcv.diag() added for completeness.
1.3.5
-Fixed small bug in compare.kda.cv() and compare.kda.diag.cv().
1.3.4
-RGL-type plots added for 3-d data. Specification of 3-d contour
levels now same order as 2-d contours.
1.3.x
-Multivariate (for 3 to 6 dimensions inclusive) bandwidth selectors
added for Hpi(), Hpi.diag(), Hlscv(), Hlscv.diag() and Hscv(). NB:
because Hbcv() and Hbcv.diag() performed poorly for 2-d, these
weren't implemented in higher dimensions.
1.2.x
-Package checked by CRAN testers and accepted on the CRAN website. To
pass all the necessary checks involved some internal programming
changes but has not affected the user interface.
-The child mortality data set unicef is used in the examples.
1.1.x
-S3 type objects have been introduced. The output from kde() are
`kde' objects. The output from kda.kde() and pda.pde() are `dade'
objects. Corresponding plot functions are called automatically by
invoking `plot'.
-Kernel discriminant analysers are now available. Parametric (linear
and quadratic) discriminant analysers are accessed using `pda'.
-adapt library is no longer required. This was formerly used on the
functions for integrated squared error computations ise.mixt() and
iset.mixt().