The R Project SVN R

                Dear Emacs, please make this -*-Text-*- mode!
    **************************************************
    *                        *
    *          1.9 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 1.9.1 Patched


INSTALLATION ISSUES

    o   Installation will now work even in Norwegian and Danish locales
    which sort AA at the end (for package stats4 which has AAA.R).


BUG FIXES

    o   Various memory leaks have been plugged and uses of strcpy()
    with overlapping src and dest corrected.

    o   R CMD INSTALL now also works for /bin/sh's such as the one from
    Solaris 8 which fail when a function has the same name as a variable.

    o   The Date method for trunc() failed.

    o   window() failed if both start and end were outside the time
    range of the original series (possible if extend = TRUE).

    o   coplot(..) doesn't give an extraneous warning anymore when called
    on a fresh device.

    o   hasArg() used wrong logic to get the parent function. (sys.function()
    behaves differently from what is documented.

    o   prompt(f) now gives proper \usage{..} for
    f <- function(x, g = function(u) { x <- u^2 ; sin(x)/x }) { g(x) }

    o   package.skeleton() now uses the supplied name in the DESCRIPTION
    file.

    o   options(list('digits', 'scipen')) no longer seg.faults (PR#7078).

    o   summary.Date() now has a sensible default for 'digits'.



        CHANGES IN R VERSION 1.9.1


NEW FEATURES

    o   as.Date() now has a method for "POSIXlt" objects.

    o   mean() has a method for "difftime" objects and so summary()
    works for such objects.

    o   legend() has a new argument 'pt.cex'.

    o   plot.ts() has more arguments, particularly 'yax.flip'.

    o   heatmap() has a new 'keep.dendro' argument.

    o   The default barplot method now handles vectors and 1-d arrays
    (e.g., obtained by table()) the same, and uses grey instead of
    heat color palettes in these cases.  (Also fixes PR#6776.)

    o   nls() now looks for variables and functions in its formula in
    the environment of the formula before the search path, in the
    same way lm() etc look for variables in their formulae.


INSTALLATION ISSUES

    o   src/modules/X11/dataentry.c would not build on some XFree
    4.4.0 systems.  (This is a bug in their header files but we have
    added a workaround.)

    o   Building with gcc/g77 3.4.0 on ix86 platforms failed to produce
    a working build: the critical LAPACK routines are now compiled
    with -ffloat-store.

    o   Added patches to enable 64-bit builds on AIX 5.1: see the R-admin
    manual for details.

    o   Added some patches to allow non-IEEE-754 installations to work
    reasonably well.  (Infs and NAs are still not handled properly
    in complex arithmetic and functions such as sin().  See also
    Deprecated, as support for non-IEEE-754 installations is about
    to be removed.)

    o   Installation will now work in Estonian (et_EE*) locales, which
    sort z before u.  (PR#6958)


DEPRECATED & DEFUNCT

    o   Support for non-IEEE-754 arithmetic (which has been untested
    for some time) will be removed in the next full release.

    o   Direct use of R INSTALL|REMOVE|BATCH|COMPILE|SHLIB is
    deprecated: use R CMD instead.

    o   The gnome/GNOME graphics device is deprecated and will be
    removed in the next full release.


BUG FIXES

    o   pbinom(q, N, prob) is now more accurate when prob is close to 0.
    (PR#6757)

    o   pcauchy(x, .., log.p) is now more accurate for large x,
    particularly when log.p = TRUE. (PR#6756)

    o   pgeom(q, prob, lower.tail, log.p) is now (sometimes much) more
    accurate when prob is very small. (PR#6792)

    The code for pgeom(prob=1) assumed IEEE 754 arithmetic, and
    gave NaNs under gcc 3.4.0 -fPIC, for example.

    o   makeARIMA() was not handling an ARMA(0, 0) model correctly.

    o   as.Date() was failing on factors.  (PR#6779)

    o   min(), max() and range() were failing on "difftime" objects.

    o   as.data.frame.list() could fail on some unusual list names.
    (PR#6782)

    o   type.convert() ignored na.strings when no conversion was done.
    (PR#6781, not needed for its primary use in read.table.)

    o   Fixed a clipping problem in the quartz() device.

    o   Subsetting a factor swapped the order of the attributes, which
    identical() cares about.  (PR#6799)

    o   The L-BFGS-B option of optim() apparently needs part of its
    workspace zeroed.  (PR#6720)

    o   extractAIC.survreg() needed updating.

    o   When using the header Rmath.h in standalone mode, the case where
    TRUE, FALSE are already defined is now handled correctly.

    o   Package utils now exports several functions that are needed for
    writing Sweave drivers.

    o   Comparison of two lists/expressions was giving nonsensical
    (and often random) answers, and is now an error.

    o   The C-level function ncols was returning a random answer
    (often 0) for a 1D array.  This caused model.matrix to
    misbehave (perhaps segfault) if a term was a 1D array.  (PR#6838)

    o   The configure script now finds the pdf viewers ggv and gpdf.

    o   Workaround for the problems strptime on MacOS X has with dates
    before 1900.

    o   'R CMD build' works in a directory whose path contains spaces.
    (PR#6830 under Unix/Linux: it already worked under Windows.)
    Also 'R CMD check'.

    o   mosaicplot() stops cleanly if given a table containing missing values.

    o   install.packages() from a local CRAN was broken.

    o   bxp() fixed for e.g.,  boxplot(..., border=2:4)

    o   approx(list(x=rep(NaN,9), y=1:9), xout=NaN) does not seg.fault
    anymore (PR#6809).

    o   plot(1, pch=NA) does not give an error anymore and
    plot(1:2, pch=c("o",NA)) only prints one symbol  (PR#6876).

    o   diffinv(matrix(3, 7,0)) now works.

    o   plot.ts(z) for multivariate 'z' now properly draws all 'nc' xlab`s
    when nc > 1 and obeys 'ann=FALSE' or 'axes=FALSE'.

    o   aggregate(.data.frame) failed if the answer would have had one
    row.

    o   recordPlot() and replayPlot() failed to duplicate the display
    list, so further plotting altered the saved or replayed object.

    o   Assignments of the form adf[i,j] <- value now accept a
    data-frame value as well as a list value.

    o   dir.create() sometimes erroneously continued to report a directory
    already existed after the first instance.  (PR#6892)

    o   arima.sim() allows a null model.

    o   which.min() & which.max()'s C code now PROTECT()'s its result.

    o   Building standalone nmath did not support some of the DEBUG options.

    o   mle() got confused if start value list was not in same order as
        arguments of likelihood function (reported by Ben Bolker)

    o   backsolve(r, x, k) now allows k < nrow(x) - as its documentation
    always claimed.

    o   update.packages("mgcv") and old.packages(*) now give a better error
    message; and installed.packages("mgcv") properly returns <empty>.

    o   stats:::as.dendrogram.hclust() is documented and no longer re-sorts
    the two children at each node.  This fixes as.dendrogram(hh) for
    the case where hh is a "reordered" hclust object.

    plot.dendrogram(x) now draws leaves 'x' more sensibly.

    reorder.dendrogram() now results in a dendrogram with correct
    "midpoint"s, and hence reordered dendrograms are plotted correctly.

    stats:::midcache.dendrogram() and hence the reorder() and rev()
    dendrogram methods do not return bloated dendrograms.

    o   heatmap(*, labRow=., labCol=.) now also reorders the labels when
    specified---not only when using default labels.

    o   Copying lattice (grid) output to another device now works again
        (There were intermittent problems in 1.9.0 - PR#6915, #6947/8.)

    o   hist() uses a more robust choice of its 'diddle' factor, used
    to detect if an observation is on a bin boundary.  (PR#6931)

    o   jitter(x) now returns x when length(x) == 0.

    o   Under some rare circumstances the locale-specific tables used by
    the perl=TRUE option to grep() etc were being corrupted and so
    matches were missed.

    o   qbinom(*, prob = 0, lower.tail = FALSE) now properly gives 0.
    (PR#6972)

    o   Class "octmode" needed a "[" method to preserve the class: see
    example(file.info) for an example.



        CHANGES IN R VERSION 1.9.0


USER-VISIBLE CHANGES

    o   Underscore '_' is now allowed in syntactically valid names, and
    make.names() no longer changes underscores.  Very old code
    that makes use of underscore for assignment may now give
    confusing error messages.

    o   Package 'base' has been split into packages 'base', 'graphics',
    'stats' and 'utils'.  All four are loaded in a default
    installation, but the separation allows a 'lean and mean'
    version of R to be used for tasks such as building indices.

    Packages ctest, eda, modreg, mva, nls, stepfun and ts have been
    merged into stats, and lqs has been returned to MASS.  In all
    cases a stub has been left that will issue a warning and ensure
    that the appropriate new home is loaded.  All the time series
    datasets have been moved to package stats.  Sweave has been
    moved to utils.

    Package mle has been moved to stats4 which will become the
    central place for statistical S4 classes and methods
    distributed with base R.  Package mle remains as a stub.

    Users may notice that code in .Rprofile is run with only the
    new base loaded and so functions may now not be found.  For
    example, ps.options(horizontal = TRUE) should be preceded by
    library(graphics) or called as graphics::ps.options or,
    better, set as a hook -- see ?setHook.

    o   There has been a concerted effort to speed up the startup of
    an R session: it now takes about 2/3rds of the time of 1.8.1.

    o   A warning is issued at startup in a UTF-8 locale, as currently R
    only supports single-byte encodings.


NEW FEATURES

    o   $, $<-, [[, [[<- can be applied to environments. Only character
    arguments are allowed and no partial matching is done.  The
    semantics are basically that of get/assign to the environment with
    inherits=FALSE.

    o   There are now print() and [ methods for "acf" objects.

    o   aov() will now handle singular Error() models, with a warning.

    o   arima() allows models with no free parameters to be fitted (to
    find log-likelihood and AIC values, thanks to Rob Hyndman).

    o   array() and matrix() now allow 0-length `data' arguments for
    compatibility with S.

    o   as.data.frame() now has a method for arrays.

    o   as.matrix.data.frame() now coerces an all-logical data frame
    to a logical matrix.

    o   New function assignInNamespace() parallelling fixInNamespace.

    o   There is a new function contourLines() to produce contour
    lines (but not draw anything).  This makes the CRAN package
    clines (with its clines() function) redundant.

    o   D(), deriv(), etc now also differentiate asin(), acos(), atan(),
    (thanks to a contribution of Kasper Kristensen).

    o   The `package' argument to data() is no longer allowed to be a
    (unquoted) name and so can be a variable name or a quoted
    character string.

    o   There is a new class "Date" to represent dates (without times)
    plus many utility functions similar to those for date-times.
    See ?Date.

    o   Deparsing (including using dump() and dput()) an integer
    vector now wraps it in as.integer() so it will be source()d
    correctly.  (Related to PR#4361.)

    o   .Deprecated() has a new argument `package' which is used in
    the warning message for non-base packages.

    o   The print() method for "difftime" objects now handles arrays.

    o   dir.create() is now an internal function (rather than a call to
    mkdir) on Unix as well as on Windows.  There is now an option
    to suppress warnings from mkdir, which may or may not have
    been wanted.

    o   dist() has a new method to calculate Minkowski distances.

    o   expand.grid() returns appropriate array dimensions and dimnames
    in the attribute "out.attrs", and this is used by the
    predict() method for loess to return a suitable array.

    o   factanal(), loess() and princomp() now explicitly check for
    numerical inputs; they might have silently coded factor
    variables in formulae.

    o   New functions factorial(x) defined as gamma(x+1) and for
    S-PLUS compatibility, lfactorial(x) defined as lgamma(x+1).

    o   findInterval(x, v) now allows +/-Inf values, and NAs in x.

    o   formula.default() now looks for a "terms" component before a
    'formula' argument in the saved call: the component will have
    `.' expanded and probably will have the original environment
    set as its environment.  And what it does is now documented.

    o   glm() arguments `etastart' and `mustart' are now evaluated via
    the model frame in the same way as `subset' and `weights'.

    o   Functions grep(), regexpr(), sub() and gsub() now coerce their
    arguments to character, rather than give an error.

    The perl=TRUE argument now uses character tables prepared for
    the locale currently in use each time it is used, rather than
    those of the C locale.

    o   New functions head() and tail() in package `utils'.
    (Based on a contribution by Patrick Burns.)

    o   legend() has a new argument 'text.col'.

    o   methods(class=) now checks for a matching generic, and so no
    longer returns methods for non-visible generics (and
    eliminates various mismatches).

    o   A new function mget() will retrieve multiple values from an
    environment.

    o   model.frame() methods, for example those for "lm" and "glm",
    pass relevant parts of ... onto the default method.  (This has
    long been documented but not done.)  The default method is now
    able to cope with model classes such as "lqs" and "ppr".

    o   nls() and ppr() have a `model' argument to allow the model frame
    to be returned as part of the fitted object.

    o   "POSIXct" objects can now have a "tzone" attribute that
    determines how they will be converted and printed.  This means
    that date-time objects which have a timezone specified will
    generally be regarded as in their original time zone.

    o   postscript() device output has been modified to work around
    rounding errors in low-precision calculations in gs >= 8.11.
    (PR#5285, which is not a bug in R.)

    It is now documented how to use other Computer Modern fonts,
    for example italic rather than slanted.

    o   ppr() now fully supports categorical explanatory variables,

    ppr() is now interruptible at suitable places in the
    underlying FORTRAN code.

    o   princomp() now warns if both `x' and `covmat' are supplied,
    and returns scores only if the centring used is known.

    o   psigamma(x, deriv=0), a new function generalizes, digamma() etc.
    All these (psigamma, digamma, trigamma,...) now also work for  x < 0.

    o   pchisq(*, ncp > 0) and hence qchisq() now work with much higher
    values of ncp; it has become much more accurate in the left tail.

    o   read.table() now allows embedded newlines in quoted fields. (PR#4555)

    o   rep.default(0-length-vector, length.out=n) now gives a vector
    of length n and not length 0, for compatibility with S.

    If both `each' and `length.out' have been specified, it now
    recycles rather than fills with NAs for S compatibility.

    If both `times' and `length.out' have been specified, `times'
    is now ignored for S compatibility.  (Previously padding with
    NAs was used.)

    The "POSIXct" and "POSIXlt" methods for rep() now pass ... on
    to the default method (as expected by PR#5818).

    o   rgb2hsv() is new, an R interface the C API function with the same name.

    o   User hooks can be set for onLoad, library, detach and
    onUnload of packages/namespaces: see ?setHook.

    o   save() default arguments can now be set using option
    "save.defaults", which is also used by save.image() if option
    "save.image.defaults" is not present.

    o   New function shQuote() to quote strings to be passed to OS shells.

    o   sink() now has a split= argument to direct output to both the
    sink and the current output connection.

    o   split.screen() now works for multiple devices at once.

    o   On some OSes (including Windows and those using glibc)
    strptime() did not validate dates correctly, so we have added
    extra code to do so.  However, this cannot correct scanning
    errors in the OS's strptime (although we have been able to
    work around these on Windows).  Some examples are now tested for
    during configuration.

    o   strsplit() now has `fixed' and `perl' arguments and
    split="" is optimized.

    o   subset() now allows a `drop' argument which is passed on to
    the indexing method for data frames.

    o   termplot() has an option to smooth the partial residuals.

    o   varimax() and promax() add class "loadings" to their loadings
    component.


    o   Model fits now add a "dataClasses" attribute to the terms, which
    can be used to check that the variables supplied for
    prediction are of the same type as those used for fitting.
    (It is currently used by predict() methods for classes "lm",
    "mlm", "glm" and "ppr", as well as methods in packages MASS,
    rpart and tree.)

    o   New command-line argument --max-ppsize allows the size of the
    pointer protection stack to be set higher than the previous
    limit of 10000.

    o   The fonts on an X11() device (also jpeg() and png() on Unix)
    can be specified by a new argument `fonts' defaulting to the
    value of a new option "X11fonts".

    o   New functions in the tools package: pkgDepends, getDepList and
    installFoundDepends.  These provide functionality for assessing
    dependencies and the availability of them (either locally or
    from on-line repositories).

    o   The parsed contents of a NAMESPACE file are now stored at
    installation and if available used to speed loading the
    package, so packages with namespaces should be reinstalled.

    o   Argument `asp' although not a graphics parameter is accepted
    in the ... of graphics functions without a warning.  It now
    works as expected in contour().

    o   Package stats4 exports S4 generics for AIC() and BIC().

    o   The Mac OS X version now produces an R framework for easier linking
    of R into other programs.  As a result, R.app is now relocatable.

    o   Added experimental support for conditionals in NAMESPACE files.

    o   Added as.list.environment to coerce environments to lists
    (efficiently).

    o   New function addmargins() in the stats package to add marginal
    summaries to tables, e.g. row and column totals.  (Based on a
    contribution by Bendix Carstensen.)

    o   dendrogam edge and node labels can now be expressions (to be
    plotted via stats:::plotNode called from plot.dendrogram).
    The diamond frames around edge labels are more nicely scaled
    horizontally.

    o   Methods defined in the methods package can now include
    default expressions for arguments.  If these arguments are
    missing in the call, the defaults in the selected method will
    override a default in the generic.  See ?setMethod.

    o   Changes to package 'grid':

    - Renamed push/pop.viewport() to push/popViewport().

    - Added upViewport(), downViewport(), and seekViewport() to
      allow creation and navigation of viewport tree
      (rather than just viewport stack).

    - Added id and id.lengths arguments to grid.polygon() to allow
      multiple polygons within single grid.polygon() call.

    - Added vpList(), vpStack(), vpTree(), and current.vpTree()
      to allow creation of viewport "bundles" that may be pushed
      at once (lists are pushed in parallel, stacks in series).

      current.vpTree() returns the current viewport tree.

    - Added vpPath() to allow specification of viewport path
      in downViewport() and seekViewport().

      See ?viewports for an example of its use.

      NOTE: it is also possible to specify a path directly,
      e.g., something like "vp1::vp2", but this is only
      advised for interactive use (in case I decide to change the
      separator :: in later versions).

    - Added "just" argument to grid.layout() to allow justification
      of layout relative to parent viewport *IF* the layout is not
      the same size as the viewport.  There's an example in
      help(grid.layout).

    - Allowed the "vp" slot in a grob to be a viewport name or a
      vpPath.  The interpretation of these new alternatives is to
      call downViewport() with the name or vpPath before drawing the
      grob and upViewport() the appropriate amount after drawing the
      grob.  Here's an example of the possible usage:

          pushViewport(viewport(w=.5, h=.5, name="A"))
          grid.rect()
          pushViewport(viewport(w=.5, h=.5, name="B"))
          grid.rect(gp=gpar(col="grey"))
          upViewport(2)
          grid.rect(vp="A", gp=gpar(fill="red"))
          grid.rect(vp=vpPath("A", "B"), gp=gpar(fill="blue"))

    - Added engine.display.list() function.  This allows the user to
      tell grid NOT to use the graphics engine display list and to handle
      ALL redraws using its own display list (including redraws after
      device resizes and copies).

      This provides a way to avoid some of the problems with resizing
      a device when you have used grid.convert(), or the gridBase package,
      or even base functions such as legend().

      There is a document discussing the use of display lists in grid
      on the grid web site
      (http://www.stat.auckland.ac.nz/~paul/grid/grid.html)

    - Changed the implementation of grob objects.  They are no longer
      implemented as external references.  They are now regular R objects
      which copy-by-value.  This means that they can be saved/loaded
      like normal R objects.  In order to retain some existing grob
      behaviour, the following changes were necessary:

      + grobs all now have a "name" slot.  The grob name is used to
        uniquely identify a "drawn" grob (i.e., a grob on the display
        list).
      + grid.edit() and grid.pack() now take a grob name as the first
        argument instead of a grob.  (Actually, they take a gPath -
        see below)
      + the "grobwidth" and "grobheight" units take either a grob
        OR a grob name (actually a gPath - see below).  Only in the
        latter case will the unit be updated if the grob "pointed to"
        is modified.

      In addition, the following features are now possible with grobs:

      + grobs now save()/load() like any normal R object.
      + many grid.*() functions now have a *Grob() counterpart.  The
        grid.*() version is used for its side-effect of drawing
        something or modifying something which has been drawn;  the
        *Grob() version is used for its return value, which is a grob.
        This makes it more convenient to just work with grob
        objects without producing any graphical output
        (by using the *Grob() functions).
      + there is a gTree object (derived from grob), which is a grob
        that can have children.  A gTree also has a "childrenvp" slot
        which is a viewport which is pushed and then "up"ed before the
        children are drawn;  this allows the children of a gTree to
        place themselves somewhere in the viewports specified in the
        childrenvp by having a vpPath in their vp slot.
      + there is a gPath object, which is essentially a concatenation
        of grob names.  This is used to specify the child of
        (a child of ...) a gTree.
      + there is a new API for creating/accessing/modifying grob objects:
        grid.add(), grid.remove(), grid.edit(), grid.get() (and their
        *Grob() counterparts can be used to add, remove, edit, or extract
        a grob or the child of a gTree.  NOTE: the new grid.edit() API
        is incompatible with the previous version.

    - Added stringWidth(), stringHeight(), grobWidth(), and grobHeight()
      convenience functions (they produce "strwidth", "strheight",
      "grobwidth", and "grobheight" unit objects, respectively).

    - Allowed viewports to turn off clipping altogether.
      Possible settings for viewport clip arg are now:

        "on"      = clip to the viewport (was TRUE)
        "inherit" = clip to whatever parent says (was FALSE)
        "off"     = turn off clipping

      Still accept logical values (and NA maps to "off")


UTILITIES

    o   R CMD check  now runs the (Rd) examples with default RNGkind
    (uniform & normal) and set.seed(1).
    example(*, setRNG = TRUE) does the same.

    o   undoc() in package `tools' has a new default of `use.values =
    NULL' which produces a warning whenever the default values of
    function arguments differ between documentation and code.
    Note that this affects "R CMD check" as well.

    o   Testing examples via massage-examples.pl (as used by R CMD
    check) now restores the search path after every help file.

    o   checkS3methods() in package 'tools' now also looks for generics
    in the loaded namespaces/packages listed in the Depends fields
    of the package's DESCRIPTION file when testing an installed
    package.

    o   The DESCRIPTION file of packages may contain a 'Suggests:'
    field for packages that are only used in examples or
    vignettes.

    o   Added an option to package.dependencies() to handle the
    'Suggests' levels of dependencies.

    o   Vignette dependencies can now be checked and obtained via
    vignetteDepends.

    o   Option 'repositories' to list URLs for package repositories
    added.

    o   package.description() has been replaced by packageDescription().

    o   R CMD INSTALL/build now skip Subversion's .svn directories as
    well as CVS directories.


C-LEVEL FACILITIES

    o   arraySubscript and vectorSubscript take a new argument which
    is a function pointer that provides access to character
    strings (such as the names vector) rather than assuming these
    are passed in.

    o   R_CheckUserInterrupt is now described in `Writing R Extensions'
    and there is a new equivalent subroutine rchkusr for calling
    from FORTRAN code.

    o   hsv2rgb and rgb2hsv are newly in the C API.

    o   Salloc and Srealloc are provided in S.h as wrappers for S_alloc
    and S_realloc, since current S versions use these forms.

    o   The type used for vector lengths is now R_len_t rather than
    int, to allow for a future change.

    o   The internal header nmath/dpq.h has slightly improved macros
    R_DT_val() and R_DT_Cval(), a new R_D_LExp() and improved
    R_DT_log() and R_DT_Clog();  this improves accuracy in several
    [dpq]-functions {for "extreme" arguments}.


DEPRECATED & DEFUNCT

    o   print.coefmat() is defunct, replaced by printCoefmat().

    o   codes() and codes<-() are defunct.

    o   anovalist.lm (replaced in 1.2.0) is now defunct.

    o   glm.fit.null(), lm.fit.null() and lm.wfit.null() are defunct.

    o   print.atomic() is defunct.

    o   The command-line arguments --nsize and --vsize are no longer
    recognized as synonyms for --min-nsize and --min-vsize (which
    replaced them in 1.2.0).

    o   Unnecessary methods coef.{g}lm and fitted.{g}lm have been
    removed: they were each identical to the default method.

    o   La.eigen() is deprecated now eigen() uses LAPACK by default.

    o   tetragamma() and pentagamma() are deprecated, since they are
    equivalent to psigamma(, deriv=2) and psigamma(, deriv=3).

    o   LTRUE/LFALSE in Rmath.h have been removed: they were
    deprecated in 1.2.0.

    o   package.contents() and package.description() have been deprecated.


INSTALLATION CHANGES

    o   The defaults for configure are now --without-zlib
    --without-bzlib --without-pcre.

    The included PCRE sources have been updated to version 4.5 and
    PCRE >= 4.0 is now required if --with-pcre is used.

    The included zlib sources have been updated to 1.2.1, and this
    is now required if --with-zlib is used.

    o   configure no longer lists bzip2 and PCRE as `additional
    capabilities' as all builds of R have had them since 1.7.0.

    o   --with-blas=goto to use K. Goto's optimized BLAS will now work.


BUG FIXES

    o   When lm.{w}fit() disregarded arguments in ... they reported
    the values and not the names.

    o   lm(singular.ok = FALSE) was looking for 0 rank, not rank < p.

    o   The substitution code for strptime in the sources no longer
    follows glibc in silently `correcting' invalid inputs.

    o   The cor() function did not remove missing values in the
    non-Pearson case.

    o   [l]choose() use a more accurate formula which also slightly
    improves p- and qhyper(); choose(n, k) now returns 0 instead
    of NaN for k < 0 or > n.

    o   find(simple.words=TRUE) (the default) was still using regular
    expressions for e.g. "+" and "*".  Also, it checked the mode
    only of the first object matching a regular expression found
    in a package.

    o   Memory leaks in [dpq]wilcox and [dqr]signrank have been plugged.
    These only occurred when multiple values of m or n > 50 were
    used in a single call. (PR#5314, plus another potential leak.)

    o   Non-finite input values to eigen(), La.eigen(), svd() and
    La.svd() are now errors: they often caused infinite
    looping.  (PR#5406, PR#4366, PR#3723: the fix for 3723/4366
    returned a vector of NAs, not a matrix, for the eigenvectors.)

    o   stepfun(x,y) now gives an error when `x' has length 0 instead
    of an invalid result (that could lead to a segmentation
    fault).

    o   buildVignettes() uses file.remove() instead of unlink() to
    remove temporary files.

    o   methods(class = "lqs") does not produce extraneous entries anymore.

    o   Directly calling a method that uses NextMethod() no longer
    produces the erroneous error message 'function is not a
    closure'.

    o   chisq.test(x, simulate.p.value = TRUE) could hang in an infinite
    loop or segfault, as r2dtable() did, when the entries in x where
    large. (PR#5701)

    o   fisher.test(x) could give a P-value of 'Inf' in similar cases which
    now result in an error (PR#4688).  It silently truncated
    non-integer 'x' instead of rounding.

    o   cutree(a, h=h) silently gave wrong results when 'a' was an
    agnes object; now gives an error and reminds of as.hclust().

    o   postscript() could crash if given a font value outside the
    valid range 1...5.

    o   qchisq(1-e, .., ncp=.) did not terminate for small e.
    (PR#6421 (PR#875))

    o   contrasts() turns a logical variable into a factor.  This now
    always has levels c("FALSE", "TRUE") even if only one (or
    none) of these occur in the variable.

    o   model.frame()'s lm and glm methods had 'data' and 'na.action'
    arguments which they ignored and have been removed.

    o   The defaults data=list() in lm() and glm() could never be
    used and have been removed.  glm had na.action=na.fail, again
    never used.

    o   tools:::.getInternalS3generics() was omitting all the members
    of the S3 group generics, which also accept methods for members.

    o   Some BLASes were returning NA %*% 0 as 0 and some as NA.  Now
    slower but more careful code is used if NAs are present. (PR#4582)

    o   package.skeleton() no longer generates invalid filenames for
    code and help files.  Also, care is taken not to generate
    filenames that differ only by case.

    o   pairs() now respects axis graphical parameters such as
    cex.main, font.main and las.

    o   Saving images of packages with namespaces (such as mle) was
    not compressing the image.

    o   When formula.default() returned a terms object, it returned a
    result of class c("terms", "formula") with different
    subsetting rules from an object of class "formula".

    o   The standalone Rmath library did not build correctly on systems
    with inaccurate log1p.

    o   Specifying asp is now respected in calls like plot(1, 10, asp=1)
    with zero range on both axes.

    o   outer() called rep() with an argument the generic does not
    have, and discarded the class of the answer.

    o   object.size() now returns a real (not integer) answer and so
    can cope with objects occupying more than 2Gb.

    o   Lookups base:: and ::: were not confining their search to the
    named package/namespace.

    o   qbinom() was returning NaN for prob = 0 or 1 or size = 0 even
    though the result is well-defined.  (In part, PR#5900.)

    o   par(mgp)[2] was being interpreted as relative to par(mgp)[3].
    (PR#6045)

    o   Versioned install was broken both with and without namespaces:
    no R code was loaded.

    o   methods(), getS3method() and the registration of S3 methods in
    namespaces were broken if the S3 generic was converted into an
    S4 generic by setting an S4 method.

    o   Title and copyright holder of the reference manual are now in
    sync with the citation() command.

    o   The validation code for POSIXlt dates and hence
    seq(, by="DSTdays") now works for large mday values (not
    just those in -1000...1000).  (PR#6212)

    o   The print() method for data frames now copes with data frames
    containing arrays (other than matrices).

    o   texi2dvi() and buildVignettes() use clean=FALSE as default
    because the option is not supported on some Solaris
    machines. For buildVignettes() this makes no difference as it
    uses an internal cleanup mechanism.

    o   The biplot() method for "prcomp" was not registered nor exported.
    (PR#6425)

    o   Latex conversion of .Rd files was missing newline before
    \end{Section} etc which occasionally gave problems, as fixed for
    some other \end{Foo} in 1.8.1.  (PR#5645)

    o   Work around a glibc bug to make the %Z format usable in strftime().

    o   The glm method for rstandard() was wrongly scaled for cases where
    summary(model)$dispersion != 1.

    o   Calling princomp() with a covariance matrix (rather than a
    list) failed to predict scores rather than predict NA as
    intended.  (PR#6452)

    o   termplot() is more tolerant of variables not in the data= argument.
    (PR#6327)

    o   isoreg() could segfault on monotone input sequences.  (PR#6494)

    o   Rdconv detected \link{\url{}} only very slowly.  (PR#6496)

    o   aov() with Error() term and no intercept incorrectly assigned
    terms to strata. (PR#6510)

    o   ftable() incorrectly handled arguments named "x".  (PR#6541)

    o   vector(), matrix(), array() and their internal equivalents
    report correctly that the number of elements specified was too
    large (rather than reporting it as negative).

    o   Minor copy-paste error in example(names).  (PR#6594)

    o   length<-() now works correctly on factors (and is now generic
    with a method for factors).

    o   x <- 2^32; x:(x+3) no longer generates an error (but gives a
    result of type "double").

    o   pgamma(30, 100, lower=FALSE, log=TRUE) is not quite 0, now.
    pgamma(x, alph) now only uses a normal approximation for
    alph > 1e5 instead of alph > 1000.  This also improves the accuracy
    of ppois().

    o   qgamma() now does one or more final Newton steps, increasing
    accuracy from around 2e-8 to 3e-16 in some cases.  (PR#2214).
    It allows values p close to 1 not returning Inf, with accuracy for
    'lower=FALSE', and values close to 0 not returning 0 for
    'log=TRUE'.  These also apply to qchisq(), e.g.,
    qchisq(1e-13, 4, lower=FALSE) is now finite and
    qchisq(1e-101, 1) is positive.

    o   gamma(-n) now gives NaN for all negative integers -n.

    o   The Unix version of browseURL() now protects the URL from the
    shell, for example allowing & and $ to occur in the URL.

    It was incorrectly attempting to use -remote "openURL()" for
    unknown browsers.

    o   extractAIC.coxph() works around an inconsistency in the
    $loglik output from coxph.  (PR#6646)

    o   stem() was running into integer overflows with nearly-constant
    inputs, and scaling badly for constant ones.  (Partly PR#6645)

    o   system() under Unix was losing the 8095th char if the output
    was split.  (PR#6624)

    o   plot.lm() gave incorrect results if there were zero weights.
    (PR#6640)

    o   Binary operators warned for inconsistent lengths on vector op
    vector operations, but not on vector op matrix ones.  (PR#6633
    and more.)

    Comparison operators did not warn about inconsistent lengths
    for real vectors, but did for integer, logical and character
    vectors.

    o   spec.pgram(x, ..., pad, fast, ...) computed the periodogram with
    a bias (downward) whenever 'pad > 0' (non-default) or 'fast = TRUE'
    (default) and nextn(n) > n where n = length(x); similarly for
    'df' (approximate degrees of freedom for chisq).

    o   dgamma(0, a) now gives Inf for a < 1 (instead of NaN), and
    so does dchisq(0, 2*a, ncp).

    o   pcauchy() is now correct in the extreme tails.

    o   file.copy() did not check that any existing `from' file had
    been truncated before appending the new contents.

    o   The QC files now check that their file operations succeeded.

    o   replicate() worked by making the supplied expression the
    body of an anonymous function(x), leading to a variable
    capture issue. Now, function(...) is used instead.

    o   chisq.test(simulate.p.value = TRUE) was returning slightly
    incorrect p values, notably p = 0 when the data gave the most
    extreme value.

    o   terms.formula(simplify = TRUE) was losing offset terms.
    Multiple offset terms were not being removed correctly if two
    of them appeared first or last in the formula.  (PR#6656)

    o   Rd conversion to latex did not add a new line before
    \end{Section} in more cases than were corrected in 1.8.1.

    o   split.default() dropped NA levels in its internal code but
    returned them as NA in all components in the interpreted code
    for factors.  (PR#6672)

    o   points.formula() had problems if there was a subset argument
    and no data argument.  (PR#6652)

    o   as.dist() does a bit more checking of its first argument and now
    warns when applied to non-square matrices.

    o   mle() gives a more understandable error message when its 'start'
    argument is not ok.

    o   All uses of dir.create() check the return value.
    download.packages() checks that destdir exists and is a directory.

    o   Methods dispatch corrects an error that failed to find
    methods for classes that extend sealed classes (class unions
    that contain basic classes, e.g.).

    o   Sweave no longer wraps the output of code chunks with
    echo=false and results=tex in Schunk environments.

    o   termplot() handles models with missing data better, especially
    with na.action=na.exclude.

    o   1:2 * 1e-100  now prints with correct number of spaces.

    o   Negative subscripts that were out of range or NA were not handled
    correctly.  Mixing negative and NA subscripts is now caught as
    an error: it was not caught on some platforms and segfaulted
    on others.

    o   gzfile() connections had trouble at EOF when used on uncompressed
    file.

    o   The Unix version of dataentry segfaulted if the `Copy' button
    was used.  (PR#6605)

    o   unlist on lists containing expressions now works (PR#5628)

    o   D(), deriv() and deriv3() now also can deal with gamma and lgamma.

    o   The X11 module can now be built against XFree86 4.4.0 headers (still
    with some warnings).

    o   seq.POSIXt(from, to, by="DSTdays") was shorter than expected
    for rare times in the UK time zone.  (PR#4558)

    o   c/rbind() did not support vectors/matrices of mode "list".  (PR#6702)

    o   summary() methods for POSIX[cl]t and Date classes coerced the
    number of NAs to a date on printing.

    o   KalmanSmooth would sometimes return NA values with NA inputs.
    (PR#6738)

    o   fligner.test() worked correctly only if data were already sorted
    by group levels. (PR#6739)



    **************************************************
    *                        *
    *          1.8 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 1.8.1


NEW FEATURES

    o   There is now a "Complex" S3 group generic (a side-effect of
    fixing up the corresponding S4 group generic).

    o   help("regex") now gives a description of the regular expressions
    used in R.

    o   The startup message now shows the R Foundation as copyright
    holder, and includes the R ISBN number and a pointer to the new
    citation() function.

    o   The solve() function now uses the `tol' argument for all
    non-complex cases.  The default tolerance for LINPACK is 1e-7,
    as before.  For LAPACK it currently is .Machine$double.eps but
    may be changed in later versions of R.

    o   help.search() now defaults to agrep = FALSE when keyword= is
    specified, since no one wants fuzzy matching of categories.

    o   Function texi2dvi() in package tools can be used to compile
    latex files from within R, provided the OS has a command
    texi2dvi or texify.

    o   Objects with formal S4 classes saved in pre-1.8 versions and
    loaded into the current version have incompatible class
    attributes (no package information).  A new function,
    fixPre1.8() in package methods, will fix the class attributes.
    See the help for this function.

    o   heatmap() allows Rowv/Colv = NA, suppressing the corresponding
    dendrogram.

    o   An "antifeature": Tcl 8.0 is now officially unsupported. In 1.8.0
    it just didn't work. This very old version lacks several features
    that are needed for the new version of the tcltk package. R will
    still build the tcltk package against Tcl 8.0 but the resulting
    package will not load.


BUG FIXES

    o   symnum(x) now behaves as documented when length(x) == 0 and uses
    lower.triangular = FALSE for logical arrays.

    o   c() now has a method for "noquote" objects and hence works as
    expected.

    o   split(1:10, c(1,2)) no longer gives a spurious warning.

    o   The "Complex" S4 group generic now works.

    o   abbreviate() doesn't get into infinite loops on input that differs
    only by leading/trailing space

    o   Added check for user interrupt in Rvprintf to allow printing to be
    interrupted.

    o   Fixed bug that would cause segfault on protect stack overflow.

    o   crossprod() on matrices with zero extents would return an
    uninitialized matrix (rather than one filled with zeros).

    o   DF[[i,j]] for a data frame used row j and col i, not as intended
    row i and col j.

    o   Even more user errors in recursive indexing of lists are now
    caught.  (PR#4486)

    o   cor(<matrix>, use = "pairwise") gave wrong result in 1.8.0 (only).
    (PR#4646)

    o   merge.data.frame() could give incorrect names when one of the
    arguments had only one column.  (PR#4299)

    o   Subsetting a one-dimensional array dropped dimensions even when
    they were not of length one.  (Related to PR#4110)

    o   The plot() method for `ecdf' objects, plot.ecdf(), now allows to
    set a `ylab' argument (different from the default).

    o   cor.test(*, method="spearman") gave wrong results `randomly'
    (because of a wrong type passed to C; PR#4718).

    o   dist() objects with NA's didn't print these, now do. (PR#4866).

    o   regexpr(fixed = TRUE) returned 0-based indices.

    o   df[, length_1_index] <- value did not recycle short rhs. (PR#4820)

    o   median() no longer `works' for odd-length factor variables.

    o   packageStatus() is more robust to failing to contact a
    repository, and contacts the correct paths in the repositories
    under Windows.

    o   .setOldIs (methods) contained a typo stopping POSIXct objects (etc)
    being included in formal classes.

    o   terms() sometimes removed offset() terms incorrectly, since it
    counted by variables and not terms.  Its "offset" attribute
    was incorrectly documented as referring to terms not
    variables. (Related to PR#4941)

    o   buildVignettes() and pkgVignettes() in package tools are now
    documented.  The call to texi2dvi is wrapped in the new
    function texi2dvi() which also works on Windows.

    o   hclust() was sometimes not finding the correct inter-cluster
    distances with non-monotone methods.  (PR#4195)

    o   plot.hclust() now tolerates mode changes on dumped objects. (PR#4361)

    o   prompt() no longer insists files are in the current directory.
    (PR#4978)

    o   filter() did not use init in reverse order as documented. (PR#5017)

    o   contrasts<-() and model.matrix() now have sanity checks that
    factors having at least 2 levels (or one level and a contrast
    matrix): model.matrix() gave nonsensical results for 0-level
    factors.

    o   writeChar() could segfault if more characters were requested
    than exist. (PR#5090)

    o   round() and signif() dropped attributes with 0-length inputs,
    only. (PR#4710)

    o   The default graphics device in the GNOME interface was gtk,
    which is no longer in the base package. It is now X11.

    o   The print button on the toolbar of the GNOME graphics device
    did not work.

    o   The example code on the man page for TkWidgetcmds had not been
    updated after the change that made tkget (et al.) return
    tclObj objects, so the "Submit" button didn't work.

    o   Rd conversion to latex did not add a new line before
    \end{Section} for the section environments, which caused
    problems if the last thing in a section was \preformatted{}
    (and potentially elsewhere).

    o   Under some circumstances mosaicplot() failed if main was
    supplied as it was passed on to model.frame.default().

    o   Conversion to POSIXlt (including printing) of POSIXct dates
    before 1902 and after 2038 computed which were leap years from
    (year-1900) so got some xx00 years wrong.


        CHANGES IN R VERSION 1.8.0


MACOS CHANGES

    o   As from this release there is only one R port for the
    Macintosh, which runs only on Mac OS X. (The `Carbon' port has
    been discontinued, and the `Darwin' port is part of the new
    version.)   The current version can be run either as a
    command-line application or as an `Aqua' console.  There is a
    `Quartz' device quartz(), and the download and installation of
    both source and binary packages is supported from the Aqua
    console.  Those CRAN and BioC packages which build under Mac OS X
    have binary versions updated daily.


USER-VISIBLE CHANGES

    o   The defaults for glm.control(epsilon=1e-8, maxit=25) have been
    tightened: this will produce more accurate results, slightly
    slower.

    o   sub, gsub, grep, regexpr, chartr, tolower, toupper, substr,
    substring, abbreviate and strsplit now handle missing values
    differently from "NA".

    o   Saving data containing name space references no longer warns
    about name spaces possibly being unavailable on load.

    o   On Unix-like systems interrupt signals now set a flag that is
    checked periodically rather than calling longjmp from the
    signal handler.  This is analogous to the behavior on Windows.
    This reduces responsiveness to interrupts but prevents bugs
    caused by interrupting computations in a way that leaves the
    system in an inconsistent state.  It also reduces the number
    of system calls, which can speed up computations on some
    platforms and make R more usable with systems like Mosix.


CHANGES TO THE LANGUAGE

    o   Error and warning handling has been modified to incorporate a
    flexible condition handling mechanism.  See the online
    documentation of 'tryCatch' and 'signalCondition'.  Code that
    does not use these new facilities should remain unaffected.

    o   A triple colon operator can be used to access values of internal
    variables in a name space (i.e. a:::b is the value of the internal
    variable b in name space a).

    o   Non-syntactic variable names can now be specified by inclusion
    between backticks `Like This`.  The deparse() code has been
    changed to output non-syntactical names with this convention,
    when they occur as operands in expressions.  This is controlled
    by a `backtick' argument, which is by default TRUE for
    composite expressions and FALSE for single symbols.  This
    should give minimal interference with existing code.

    o   Variables in formulae can be quoted by backticks, and such
    formulae can be used in the common model-fitting functions.
    terms.formula() will quote (by backticks) non-syntactic names
    in its "term.labels" attribute.  [Note that other code using
    terms objects may expect syntactic names and/or not accept
    quoted names: such code will still work if the new feature is
    not used.]


NEW FEATURES

    o   New function bquote() does partial substitution like LISP backquote.

    o   capture.output() takes arbitrary connections for `file' argument.

    o   contr.poly() has a new `scores' argument to use as the base set
    for the polynomials.

    o   cor() has a new argument `method = c("pearson","spearman","kendall")'
    as cor.test() did forever. The two rank based measures do work with
    all three missing value strategies.

    o   New utility function cov2cor() {Cov -> Corr matrix}.

    o   cut.POSIXt() now allows `breaks' to be more general intervals
    as allowed for the `by' argument to seq.POSIXt().

    o   data() now has an 'envir' argument.

    o   det() uses an LU decomposition and LAPACK.  The `method'
    argument to det() no longer has any effect.

    o   dev.control() now accepts "enable" as well as "inhibit".
    (Wishlist PR#3424)

    o   *, - and / work more generally on "difftime" objects, which now
    have a diff() method.

    o   dt(*, ncp = V)  is now implemented, thanks to Claus Ekstroem.

    o   dump() only quotes object names in the file where necessary.

    o   eval() of a promise forces the promise

    o   file.path() now returns an empty character vector if given at
    least one zero-length argument.

    o   format() and hence print() make an effort to handle corrupt
    data frames, with a warning.

    o   format.info() now also works with `nsmall' in analogy with
    format.default().

    o   gamma(n) is very slightly more precise for integer n in 11:50.

    o   ? and help() will accept more un-quoted arguments, e.g. NULL.

    o   The "?" operator has new forms for querying documentation on
    S4 methods.  See the online documentation.

    o   New argument frame.plot = axes (== TRUE)  for filled.contour().

    o   New argument fixed = TRUE for grep() and regexpr() to avoid the
    need to escape strings to match.

    o   grep(x, ..., value = TRUE) preserves names of x.

    o   hist.POSIXt() can now pass arguments to hist.default()

    o   legend() and symbols() now make use of xy.coords() and accept
    a wider range of coordinate specifications.

    o   Added function library.dynam.unload() to call dyn.unload() on
    a loaded DLL and tidy up.  This is called for all the standard
    packages in namespaces with DLLs if their namespaces are unloaded.

    o   lm(singular.ok = FALSE) is now implemented.

    o   Empty lm() and glm() fits are now handled by the normal
    code: there are no methods for classes "lm.null" and
    "glm.null".  Zero-rank fits are handled consistently.

    o   make.names() has improvements, and there is a new auxiliary
    function make.unique().  (Based on code contributed by Tom
    Minka, since converted to a .Internal function.)  In
    particular make.names() now recognises that names beginning
    with a dot are valid and that reserved words are not.

    o   methods() has a print method which asterisks functions which
    are not user-visible.  methods(class = "foo") now lists
    non-visible functions, and checks that there is a matching generic.

    o   model.matrix() now warns when it removes the response from the
    rhs of the formula: that this happens is now documented on its
    help page.

    o   New option `locatorBell' to control the confirmation beep
    during the use of locator() and identify().

    o   New option("scipen") provides some user control over the
    printing of numbers in fixed-point or exponential notation.
    (Contributed by David Brahm.)

    o   plot.formula() now accepts horizontal=TRUE and works correctly
    when boxplots are produced.  (Wishlist PR#1207)  The code has
    been much simplified and corrected.

    o   polygon() and rect() now interpret density < 0 or NA to mean
    filling (by colour) is desired: this allows filling and
    shading to be mixed in one call, e.g. from legend().

    o   The predict() methods for classes lm, glm, mlm and lqs take a
    `na.action' argument that controls how missing values in
    `newdata' are handled (and defaults to predicting NA).
    [Previously the value of getOption("na.action") was used and
    this by default omitted cases with missing values, even if set
    to `na.exclude'.]

    o   print.summary.glm() now reports omitted coefficients in the
    same way as print.summary.lm(), and both show them as NAs in
    the table of coefficients.

    o   print.table() has a new argument `zero.print' and is now
    documented.

    o   rank(x, na.last = "keep") now preserves NAs in `x', and the
    argument `ties.method' allows to use non-averaging ranks in the
    presence of ties.

    o   read.table()'s 'as.is' argument can be character, naming columns
    not to be converted.

    o   rep() is now a generic function, with default, POSIXct and
    POSIXlt methods.  For efficiency, the base code uses rep.int()
    rather than rep() where possible.

    o   New function replicate() for repeated evaluation of expression
    and collection of results, wrapping a common use of sapply()
    for simulation purposes.

    o   rev() is now a generic function, with default and dendrogram
    methods.

    o   serialize() and unserialize() functions are available for
    low-level serialization to connections.

    o   socketSelect() allows waiting on multiple sockets.

    o   sort(method = "quick", decreasing = TRUE) is now implemented.

    o   sort.list() has methods "quick" (a wrapper for sort(method =
    "quick", index.return = TRUE) and "radix" (a very fast method
    for small integers).  The default "shell" method works faster
    on long vectors with many ties.

    o   stripchart() now has `log', `add' and `at' arguments.

    o   strsplit(x, *) now preserves names() but won't work for
    non-character `x' anymore {formerly used as.character(x),
    destroying names(x)}.

    o   textConnection() now has a local argument for use with output
    connections.  local = TRUE means the variable containing the
    output is assigned in the frame of the caller.

    o   Using UseMethod() with more than two arguments now gives a
    warning (as R-lang.texi has long claimed it did).

    o   New function vignette() for viewing or listing vignettes.

    o   which.min(x) and which.max(x) now preserve names.

    o   xy.coords() coerces "POSIXt" objects to "POSIXct", allowing
    lines etc to added to plot.POSIXlt() plots.

    o   .Machine has a new entry, sizeof.pointer.

    o   .Random.seed is only looked for and stored in the user's
    workspace.  Previously the first place a variable of that name
    was found on the search path was used.

    o   Subscripting for data.frames has been rationalized:

    - Using a single argument now ignores any `drop' argument
      (with a warning).  Previously using `drop' inhibited list-like
      subscripting.

    - adf$name <- value now checks for the correct length of
      `value', replicating a whole number of times if needed.

    - adf[j] <- value and adf[[j]] <- value did not convert
      character vectors to factors, but adf[,j] <- value did.
      Now none do.  Nor is a list `value' coerced to a data frame
      (thereby coercing character elements to factors).

    - Where replicating the replacement value a whole number of
      times will produce the right number of values, this is
      always done (rather than some times but not others).

    - Replacement list values can include NULL elements.

    - Subsetting a data frame can no longer produce duplicate
      column names.

    - Subsetting with drop=TRUE no longer sometimes drops
      dimensions on matrix or data frame columns of the data frame.

    - Attributes are no longer stripped when replacing part of a column.

    - Columns added in replacement operations will always be
      named, using the names of a list value if appropriate.

    - as.data.frame.list() did not cope with list names such as
      `check.rows',  and formatting/printing data frames with such
      column names now works.

    - Row names in extraction are still made unique, but without
      forcing them to be syntactic names.

    - adf[x] <- list() failed if x was of length zero.


    o   Setting dimnames to a factor now coerces to character, as S
    does.  (Earlier versions of R used the internal codes.)

    o   When coercion of a list fails, a meaningful error message is given.

    o   Adding to NULL with [[ ]] generates a list if more than one
    element is added (as S does).

    o   There is a new command-line flag --args that causes the rest of
    the command line to be skipped (but recorded in commandArgs()
    for further processing).

    o   S4 generic functions and method dispatch have been modified to
    make the generic functions more self-contained (e.g., usable
    in apply-type operations) and potentially to speed dispatch.

    o   The data editor is no longer limited to 65535 rows, and will
    be substantially faster for large numbers of columns.

    o   Standalone Rmath now has a get_seed function as requested (PR#3160).

    o   GC timing is not enabled until the first call to gc.time(); it
    can be disabled by calling gc.time(FALSE).  This can speed up
    the garbage collector and reduce system calls on some
    platforms.


STANDARD PACKAGES

    o   New package 'mle'. This is a simple package to find maximum
    likelihood estimates, and perform likelihood profiling and
    approximate confidence limits based upon it.  A well-behaved
    likelihood function is assumed, and it is the responsibility
    of the user to gauge the applicability of the asymptotic
    theory.  This package is based on S4 methods and classes.

    o   Changes in package 'mva':

    - factanal() now returns the test statistic and P-value formerly
      computed in the print method.

    - heatmap() has many more arguments, partly thanks to Wolfgang
      Huber and Andy Liaw.

    - Arguments `unit' and `hmin' of plclust() are now implemented.

    - prcomp() now accepts complex matrices, and there is biplot()
      method for its output (in the real case).

    - dendrograms are slightly better documented, methods working with
      "label", not "text" attribute.  New rev() method for dendrograms.

    - plot.dendrogram() has an explicit `frame.plot' argument
      defaulting to FALSE (instead of an implicit one defaulting to TRUE).

    o   Changes in package 'tcltk':

    - The package is now in a namespace.  To remove it you will
      now need to use unloadNamespace("tcltk").

    - The interface to Tcl has been made much more efficient by
      evaluating Tcl commands via a vector of Tcl objects rather
      than by constructing the string representation.

    - An interface to Tcl arrays has been introduced.

    - as.tclObj() has gained a `drop' argument to resolve an
      ambiguity for vectors of length one.

    o   Changes in package 'tools':

    - Utilities for testing and listing files, manipulating file
      paths, and delimited pattern matching are now exported.

    - Functions checkAssignFuns(), checkDocArgs() and checkMethods()
      have been renamed to checkReplaceFuns(), checkDocFiles(), and
      checkS3methods, to given better descriptions of what they do.

    - R itself is now used for analyzing the markup in the \usage
      sections.  Hence in particular, replacement functions or S3
      replacement methods are no longer ignored.

    - checkDocFiles() now also determines 'over-documented' arguments
      which are given in the \arguments section but not in \usage.

    - checkDocStyle() and checkS3Methods() now know about internal S3
      generics and S3 group generics.

    - S4 classes and methods are included in the QC tests.
      Warnings will be issued from undoc() for classes  and
      methods defined but not documented.  Default methods
      automatically generated from nongeneric functions do not
      need to be documented.

    - New (experimental) functions codocClasses() and codocData()
      for code/documentation consistency checking for S4 classes and
      data sets.

    o   Changes in package 'ts':

    - arima.sim() now checks for inconsistent order specification
      (as requested in PR#3495: it was previously documented not to).

    - decompose() has a new argument `filter'.

    - HoltWinters() has new arguments `optim.start' and
      `optim.control', and returns more components in the fitted
      values.  The plot method allows `ylim' to be set.

    - plot.ts() has a new argument `nc' controlling the number of
      columns (with default the old behaviour for plot.mts).

    - StructTS() now allows the first value of the series to be
      missing (although it is better to omit leading NAs). (PR#3990)


USING PACKAGES

    o   library() has a pos argument, controlling where the package is
    attached (defaulting to pos=2 as before).

    o   require() now maintains a list of required packages in the
    toplevel environment (typically, .GlobalEnv).  Two features
    use this:  detach() now warns if a package is detached that is
    required by an attached package, and packages that install
    with saved images no longer need to use require() in the
    .First as well as in the main source.

    o   Packages with name spaces can now be installed using --save.

    o   Packages that use S4 classes and methods should now work with
    or without saved images (saved images are still recommended
    for efficiency), writing setMethod(), etc. calls with the
    default for argument `where'.  The topenv() function and
    sys.source() have been changed correspondingly.  See the
    online help.

    o   Users can specify in the DESCRIPTION file the collation order
    for files in the R source directory of a package.


DOCUMENTATION CHANGES

    o   Changes in R documentation format:

    - New logical markup commands for emphasizing (\strong) and
      quoting (\sQuote and \dQuote) text, for indicating the usage
      of an S4 method (\S4method), and for indicating specific kinds
      of text (\acronym, \cite, \command, \dfn, \env, \kbd, \option,
      \pkg, \samp, \var).

    - New markup \preformatted for pre-formatted blocks of text
      (like \example but within another section).  (Based on a
      contribution by Greg Warnes.)

    - New markup \concept for concept index entries for use by
      help.search().

    o   Rdconv now produces more informative output from the special
    \method{GENERIC}{CLASS} markup for indicating the usage of S3
    methods, providing the CLASS info in a comment.

    o   \dontrun sections are now marked within comments in the
    user-readable versions of the converted help pages.

    o   \dontshow is now the preferred name for \testonly.


INSTALLATION CHANGES

    o   The zlib code in the sources is used unless the external
    version found is at least version 1.1.4 (up from 1.1.3).

    o   The regression checks now have to be passed exactly, except
    those depending on recommended packages (which cannot be
    assumed to be present).

    o   The target make check-all now runs R CMD check on all the
    recommended packages (and not just runs their examples).

    o   There are new macros DYLIB_* for building dynamic libraries,
    and these are used for the dynamic Rmath library (which was
    previously built as a shared object).

    o   If a system function log1p is found, it is tested for accuracy
    and if inadequate the substitute function in src/nmath is
    used, with name remapped to Rlog1p.  (Apparently needed on
    OpenBSD/NetBSD.)


C-LEVEL FACILITIES

    o   There is a new installed header file R_ext/Parse.h which
    allows R_ParseVector to be called by those writing extensions.
    (Note that the interface is changed from that used in the
    unexported header Parse.h in earlier versions, and is not
    guaranteed to remain unchanged.)

    o   The header R_ext/Mathlib.h has been removed.  It was replaced by
    Rmath.h in R 1.2.0.

    o   PREXPR has been replaced by two macros, PREXPR for obtaining the
    expression and PRCODE for obtaining the code for use in eval.
    The macro BODY_EXPR has been added for use with closures.
    For a closure with a byte compiled body, the macro BODY_EXPR
    returns the expression that was compiled; if the body is not
    compiled then the body is returned.  This is to support byte
    compilation.

    o   Internal support for executing byte compiled code has been added.
    A compiler for producing byte compiled code will be made available
    separately and should become part of a future R release.

    o   On Unix-like systems calls to the popen() and system() C library
    functions now go through R_popen and R_system.  On Mac OS X these
    suspend SIGALRM interrupts around the library call.  (Related to
    PR#1140.)


UTILITIES

    o   R CMD check accepts "ORPHANED" as package maintainer. Package
    maintainers can now officially orphan a package, i.e., resign
    from maintaining a package.

    o   R CMD INSTALL (Unix only) is now 'safe': if the attempt to
    install a package fails, leftovers are removed.  If the package
    was already installed, the old version is restored.

    o   R CMD build excludes possible (obsolete) data and vignette
    indices in DCF format (and hence also no longer rebuilds them).

    o   R CMD check now tests whether file names are valid across file
    systems and supported operating system platforms.  There is some
    support for code/documentation consistency checking for data
    sets and S4 classes.  Replacement functions and S3 methods in
    \usage sections are no longer ignored.

    o   R CMD Rdindex has been removed.


DEPRECATED & DEFUNCT

    o   The assignment operator `_' has been removed.

    o   printNoClass() is defunct.

    o   The classic Mac OS port is no longer supported, and its files
    have been removed from the sources.

    o   The deprecated argument 'white' of parse() has been removed.

    o   Methods pacf/plot.mts() have been removed and their functionality
    incorporated into pacf.default/plot.ts().

    o   print.coefmat() is deprecated in favour of printCoefmat()
    (which is identical apart from the default for na.print which
    is changed from "" to "NA", and better handling of the 0-rank
    case where all coefficients are missing).

    o   codes() and codes<-() are deprecated, as almost all uses
    misunderstood what they actually do.

    o   The use of multi-argument return() calls is deprecated: use a
    (named) list instead.

    o   anovalist.lm (replaced in 1.2.0) is now deprecated.

    o   - and Ops methods for POSIX[cl]t objects are removed: the
    POSIXt methods have been used since 1.3.0.

    o   glm.fit.null(), lm.fit.null() and lm.wfit.null() are deprecated.

    o   Classes "lm.null" and "glm.null" are deprecated and all of their
    methods have been removed.

    o   Method weights.lm(), a copy of weights.default(), has been removed.

    o   print.atomic() is now deprecated.

    o   The back-compatibility entry point Rf_log1p in standalone
    Rmath has been removed.


BUG FIXES

    o   ARMAacf() sometimes gave too many results or failed if `lag.max'
    was used.

    o   termplot() with a subset of terms now gives correct partial residuals

    o   Functions anova.glm(), contrasts(), getS3method(), glm() and
    make.tables() were applying get() without asking for a
    function and/or not starting the search in the environment of
    the caller.

    o   as.data.frame.matrix() ignored the `row.names' argument.

    o   as.data.frame.list(optional = TRUE) was converting names, and
    hence data.frame(list(...), check.names = FALSE) was.  (PR#3280)

    o   as.dist(m) {mva} now obeys `diag=TRUE' or `upper=TRUE' in all cases.

    o   as.double(list()) etc was regarded as an error, because of a
    bug in isVectorizable.

    o   On some platforms the wday component of the result of
    as.POSIXlt() was corrupted when trying to guess the DST offset
    at dates the OS was unable to handle.

    o   ave(x, g) didn't work when `g' had unused levels.

    o   biplot.default() allows xlim and ylim to be set.  (PR#3168)

    o   bgroup with a null (.) delimiter was setting font to Greek.  (PR#3099)

    o   body() and formals() were looking for named functions in
    different places: they now both look starting at the
    environment in which they are called.  Several documentation
    errors for these functions have been corrected.

    o   boxplot() was ignoring cex.axis.  (PR#2628)

    o   cut.POSIXt() now passes on ... to cut.default(), as documented.

    o   crossprod() now works for 1d arrays with unnamed dimnames (PR#4092).

    o   data() sometimes failed with multiple files, as the paths
    variable got corrupted.

    o   data.frame() failed with a nonsensical error message if it
    grabbed row names from an argument that was subsequently
    recycled.  Now they are discarded, with a warning.

    o   data.matrix() was documented to replace factors by their
    codes, but in fact re-coded into the alphabetical ordering of
    the levels.

    o   decompose() with even frequency used an asymmetric moving
    average window.

    o   demo() was using `topic' as a regexp rather than an exact match.

    o   dotchart() now does recycle the `color' argument and better
    documents the `bg' one (PR#4343).

    o   getAnywhere() didn't not correctly check for S3 methods, when
    the generic or the class name contains a "." (PR#4275).

    o   file.copy() ignored the overwrite argument.  (PR#3529)

    o   filter(method="recursive") was unnecessarily requiring the
    time series to be longer than the filter.

    o   format(*, nsmall = m) with m > 0 now returns exponential format
    less often.

    o   get() and exists() were ignoring the `mode' argument for
    variables in base.  The error message for get() now mentions
    the mode requested if not "any".  A bug in setting the NAMED
    field in do_get was fixed.

    o   getS3method(f, cl, optional=TRUE) now returns NULL if `f' does
    not exist.

    o   HoltWinters() would segfault if only gamma was optimized, and
    not converge if gamma=0 and seasonal="mult".

    o   hyperref.cfg now contains definitions for colors it uses.

    o   identify.default() detects zero-length arguments.  (PR#4057)

    o   legend() allows shading without filling again.

    o   legend(x, y, leg) doesn't triple `leg' anymore when it is a call.

    o   Corrected many problems with 0-rank (but not necessarily empty
    model) lm() and glm() fits.

    o   lm.influence() now handles 0-rank models, and names its output
    appropriately.  It also ensures that hat values are not greater
    than one, and rounds values within rounding error of one.

    o   The `method' argument to loess() did not work.  (PR#3332)

    o   lsfit() was returning incorrect residuals for 0-rank fits.

    o   methods("$") and methods("$<-") were failing to find methods.

    o   methods() and getS3method() find methods if the generic
    dispatches on a name other than its own.  (The cases of
    coefficients() and fitted.values() were fixed in 1.7.1.)

    o   model.matrix.default() was throwing an error on 0-term models,
    but now handles them correctly.

    o   Printing `nls' objects misbehaved when `data' was a composite
    expression.

    o   .NotYetImplemented() gave "Error in .NotYet...(): .."

    o   numericDeriv() was failing if the first argument was a name
    rather than a call.  (PR#3746)

    o   pacf() was failing if called on a one-column matrix.

    o   paste() applied to 0-length vectors gave "" not a 0-length vector.

    o   The length of a string specification of par(lty=) is now checked: it
    should be 2, 4, 6 or 8.

    o   Using lty=as.integer(NA) and as.double(NA) were being accepted
    but giving nonsensical results.  Those are not documented
    valid values for lty.  (PR#3217)

    o   Erroneously calling par(new=TRUE) with no plot was not caught
    and so could result in invalid graphics files.  (PR#4037)

    o   par(tck=) was being interpreted incorrectly.  It is now
    documented in the same way as S, and now behaves as
    documented.  (PR#3504)

    o   plclust() [and hence plot.hclust()] sometimes now uses correct `ylim's
    also in unusual cases.  (PR#4197)

    o   plot.POSIX[cl]t no longer passes col, lty, lwd to axis.POSIXt.

    o   The png(), jpeg(), png() and win.metafile() devices now
    enforce the length limit on the filename.  (PR#3466)

    o   pnorm(x, 1, 0) does not give NaN anymore;
    also, pnorm(x, m, s=Inf) == lim{s -> Inf} pnorm(x,m,s).
    Similar changes for dnorm(), cf PR#1218.

    o   On some machines the internal rounding used in postscript() was
    imperfect, causing unnecessarily verbose output (-0.00 instead of
    0) and problems with make check.

    o   qqnorm()'s result now keeps NAs from its input. (PR#3750)

    o   rank() sometimes preserved and sometimes dropped names.

    o   readBin(what = "foo") didn't convert `what' to its type.  (PR#4043)

    o   reorder.dendrogram() now properly resets the "midpoint" attributes
    such that reorder()ed dendrograms now plot properly.

    o   rmultinom(1,100, c(3, 4, 2, 0,0))[3] was NA. (PR#4431)

    o   sapply() for matrix result does not return list(NULL,NULL) dimnames
    anymore.

    o   scan() now interprets quoting in fields to be skipped.  (PR#4128)

    o   seq.POSIXt(from, to, by="DSTday") was failing or calculating
    the length incorrectly.

    o   sort() and unique.default() were failing on 0-level factors.

    o   step() adds a fuzz for reduction in AIC for 0-df terms.  (PR#3491)

    o   str(x) gives better output when x is of mode "(".  Its "dendrogram"
    method obeys the `give.attr' argument which now defaults to FALSE.

    o   strwidth(f) and strheight(f) could seg.fault when `f' was a
    function.  The fix [to C-level coerceVector()] now gives an error
    instead of passing through.  This may catch other potential
    problems.

    o   Sweave() reports the chunk number rather than the driver call when
    a try error gets caught.

    o   trunc.POSIXt(x) for 0-length x does not return invalid structures
    anymore.  (PR#3763).

    o   warnings() now returns NULL instead of an error when no warnings
    have occured yet.  (PR#4389)

    o   Using write.table() setting the `dec' argument and with no
    numeric columns failed.  (PR#3532)

    o   $<- did not duplicate when it needed to.

    o   Recursive indexing of lists had too little error-checking.
    (related to PR#3324)

    o   Removed warning about names in persistent strings when a
    namespace is saved.

    o   Fixed some malformed error messages in the methods package.

    o   pipes were not opening properly when profiling on a Mac OS.
    (PR#1140)

    o   Lapack error messages (PR#3494) and call to DGEQP3 (PR#2867) are
    corrected.

    o   Rd conversion was limiting a file to 1000 pairs of braces,
    without any warning.  Now the limit is 10000, with a warning.
    (PR#3400)

    o   In the tcltk package, the tkimage.*() commands were defined
    nonsensically as widget commands. They have been redefined to be
    more useful now.

    o   Registered group generics were not being used. (PR#3536)

    o   Subsetting data frames did not always correctly detect that
    non-existent columns were specified.

    o   There are many more checks for over-running internal buffers,
    almost always reporting errors.

    o   Added some buffer overflow checking in gram.y.

    o   Internals for complex assignment did not check that function name
    was a symbol, which could cause a segfault.

    o   Fixed bug in S4 methods dispatch that made local variables in the
    generic visible when executing the body of a method, thus violating
    lexical scope.




    **************************************************
    *                        *
    *          1.7 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 1.7.1


NEW FEATURES

    o   The help pages give appropriate references to the Blue,
    White or Green books for functions based on the descriptions
    of S functions given there. (E&OE)

    o   Function getAnywhere() can find non-exported objects, for
    namespaces or registered methods.


DEPRECATED & DEFUNCT

    o   The (unimplemented) argument 'white' of parse() is deprecated.

    o   The tkfilefind demo in the tcltk library is deprecated, since
    it never worked well, and apparently not at all with Tcl/Tk 8.4.


BUG FIXES

    o   print.table() used too much white space in some cases in 1.7.0.

    o   selectMethod() failed if `f' was a non-generic and
    optional=TRUE, and gave a confusing error message if
    optional=FALSE.

    o   pchisq(*, ncp) and qchisq(*, ncp) work in more cases for large ncp
    or quantile and give warning or error messages otherwise.

    o   str(x) now also works when x is an "externalptr" (or "weakref").

    o   rbeta(), rf(), and rt() now support infinite parameter values;
    other distributions return NaN instead of NA for such.

    o   Redefining a class is now safer if the new definition
    generates an error (previously some invalid metadata could be
    left behind).

    o   A number of errors are now caught in setClass() that
    previously either went unchecked or waited until new() to
    appear:
    - classes may not contain themselves, directly or indirectly;
    - classes appearing either as slots or as superclasses must
      themselves be defined;
    - slot names (direct or inherited) must be unique.
    In related changes, prototype() now works as documented, and is the
    recommended way to provide prototype objects.

    o   Sorting an ordered factor would return an unordered one.
    This caused some trouble with panel.superpose (PR#974).

    o   methods() could return duplicates if a method in a namespace
    was both exported and registered.

    o   The internal zip.unpack() could crash if more than 500 files
    were to be extracted. (PR#2818)

    o   The "r+" and "r+b" modes of opening file connections disallowed
    writing.

    o   library() now warns the user if the chosen package name
    doesn't match the internal package name, and corrects the
    error. (PR#2816)

    o   qr(LAPACK=TRUE) (and qr for complex arguments) might have failed
    to pivot for rank-deficient inputs. (PR#2867)

    o   Only re-mapped symbols are exported by regex.o, to avoid
    problems with embedded R on RedHat 9.

    o   arima() did not set transform.pars to FALSE if AR parameters
    were fixed, although it claimed to.

    o   pnorm() was slower than necessary in the outer tails in some
    cases due to a typo in the improvements from PR#699.  (PR#2883)

    o   setGeneric() and setMethod() now catch some examples where the
    generic and the method have different argument lists; the
    evaluator checks for internal consistency of these argument lists.

    o   expand.grid(x) {the rare case of one argument} now treats factor
    levels as in the typical case of two or more arguments.

    o   Some implicit coercions to lists could cause segfaults, e.g.
      x <- matrix(nrow=20000, ncol=20); x$any <- numeric(0)
    due to a PROTECT bug. (PR#2923)

    o   The replacement functions for colnames() and rownames() did not
    work for arrays with more than two dimensions.  They could
    create dimnames of the form list(NULL, NULL) rather than
    remove the dimnames attribute.

    o   termplot() gave incorrect answers with rug=TRUE or
    partial=TRUE for factors whose levels were not in
    lexicographical order.

    o   A serious performance flaw in as() computations was fixed (the
    methods were not being cached properly.)

    o   model.frame(~1, data) always returned 1 row. (PR#2958)

    o   The data editor was truncating objects to 65535 rows.  Pro
    tem, editing objects with more than 65535 rows is an error,
    and objects cannot be extended beyond that row. This restriction
    will be removed in 1.8.0.  (PR#2962)

    o   A bug could produce apparent loops in formal method selection
    when inheritance was restricted (used for the as() function).
    A related problem sometimes broke attaching a package that had
    methods for basic functions, such as names(), used in method
    selection.

    o   Empty expressions as in return(x,) could generate subsequent
    segfaults: they are now errors.  (PR#2880)

    o   The Kinderman-Ramage Normal Random Generator had several
    problems leading to not-quite normally distributed variates
    (PR#2846).  One problem was traced to an error in the original
    1976 JASA paper!  Thanks to Josef Leydold and his team for
    investigating this. The old generator has been retained for
    reproducibility of older results, under the name
    "Buggy Kinderman-Ramage".  A warning is issued if you select it
    (also indirectly via RNGversion()).

    o   promptMethods() now puts the \alias lines for methods in the
    normal place, near the top of the file, and quotes class
     names in signatures.

    o   getS3method() and methods() were not finding methods for
    coefficients() and fitted.values() (which dispatch on "coef"
    and "fitted" respectively).

    o   scan() (and hence read.table) was not finding matches for
    separator chars with the upper bit set.  (PR#3035)

    o   lm.(w)fit failed if the fit had rank 0.

    o   lqs() did not report explicitly that it had failed if all
    samples gave singular fits.

    o   predict.lm(*, se=TRUE) {w/ weights, w/o newdata} now gives correct
    SE's.  (PR#3043)

    o   cor.test(x, y, method="spearman") now also works for
    length(x) > 1290.

    o   Matrices were printed mis-aligned if right=TRUE and na.print
    was specified.  (PR#3058)

    o   R CMD check gives now a clearer message when latex produces
    errors on the package manual.  (PR#3070)

    o   isSeekable() was incorrectly returning FALSE on all file connections.

    o   tkpager() wasn't quite using its title and header arguments in
    the way prescribed by file.show()

    o   legend(*, pch=p, lty=l) now works better when `p' or `l' have
    NAs.

    o   All braces in regular expressions used by Sweave() are now
    escaped by a backslash.

    o   unloadNamespace() failed because getNamespaceImports() now coerces a
    string argument to a name space.

    o   deriv3 gave incorrect Hessians for some very simple
    expressions such as expression(x*y) (since the comments in the
    C code were incorrect).  (PR#2577)

    o   power.t.test(..., delta=NULL,alternative='two.sided') failed. (PR#2993)

    o   Lines on postscript() plots with thousands of segments might
    have been plotted inaccurately in 1.7.0. (PR#3132)

    Solid lines in postscript() output are split into groups of 1000
    segments to help some PostScript interpreters (typically old
    level-1 interpreters).

    o   cut.POSIXt failed when the breaks were date/time objects. (PR#3181)

    o   Usage of methods in dist.Rd is now correctly documented
    (as.matrix.dist() is not an exported symbol).

    o   The predict() method for ar fits was not retrieving the series
    from the parent environment.

    o   eigen() and La.eigen() were not returning a matrix of
    eigenvectors for a 1x1 input.

    o   hsv() and rgb() now return character(0) when one of their args has
    length 0.  This also fixes terrain.color(1). (PR#3233)

    o   [[<-.data.frame checked if a replacment was too short, but not
    if it was too long.  (related to PR#3229)

    o   qt(x, df) was quite inaccurate for df=1+epsilon; it is now much more
    accurate for df in (1,2) and more precise for other df. (PR#2991)

    o   qbeta() now has slightly improved C code in two places, as suggested
    in the 2nd followup to PR#2894.



        CHANGES IN R VERSION 1.7.0


USER-VISIBLE CHANGES

    o   solve(), chol(), eigen() and svd() now use LAPACK routines
    unless a new back-compatibility option is turned on.  The
    signs and normalization of eigen/singular vectors may change
    from earlier versions.

    o   The `methods', `modreg', `mva', `nls' and `ts' packages
    are now attached by default at startup (in addition to `ctest').
    The option "defaultPackages" has been added which contains the
    initial list of packages.  See ?Startup and ?options for details.
    Note that .First() is no longer used by R itself.

    class() now always (not just when `methods' is attached) gives
    a non-null class, and UseMethod() always dispatches on the
    class that class() returns.  This means that methods like
    foo.matrix and foo.integer will be used.  Functions oldClass()
    and oldClass<-() get and set the "class" attribute as R
    without `methods' used to.

    o   The default random number generators have been changed to
    `Mersenne-Twister' and `Inversion'.  A new RNGversion()
    function allows you to restore the generators of an earlier R
    version if reproducibility is required.

    o   Namespaces can now be defined for packages other than `base':
    see `Writing R Extensions'.  This hides some internal objects
    and changes the search path from objects in a namespace.  All
    the base packages (except methods and tcltk) have namespaces,
    as well as the recommended packages `KernSmooth', `MASS',
    `boot', `class', `nnet', `rpart' and `spatial'.

    o   Formulae are not longer automatically simplified when terms()
    is called, so the formulae in results may still be in the
    original form rather than the equivalent simplified form
    (which may have reordered the terms): the results are now
    much closer to those of S.

    o   The tables for plotmath, Hershey and Japanese have been moved
    from the help pages (example(plotmath) etc) to demo(plotmath) etc.

    o   Errors and warnings are sent to stderr not stdout on
    command-line versions of R (Unix and Windows).

    o   The R_X11 module is no longer loaded until it is needed, so
    do test that x11() works in a new Unix-alike R installation.


NEW FEATURES

    o   if() and while() give a warning if called with a vector condition.

    o   Installed packages under Unix without compiled code are no
    longer stamped with the platform and can be copied to other
    Unix-alike platforms (but not to other OSes because of
    potential problems with line endings and OS-specific help files).

    o   The internal random number generators will now never return
    values of 0 or 1 for runif.  This might affect simulation
    output in extremely rare cases.  Note that this is not
    guaranteed for user-supplied random-number generators, nor
    when the standalone Rmath library is used.

    o   When assigning names to a vector, a value that is too short is
    padded by character NAs. (Wishlist part of PR#2358)

    o   It is now recommended to use the 'SystemRequirements:' field in
    the DESCRIPTION file for specifying dependencies external to the
    R system.

    o   Output text connections no longer have a line-length limit.

    o   On platforms where vsnprintf does not return the needed buffer
    size the output line-length limit for fifo(), gzfile() and
    bzfile() has been raised from 10k to 100k chars.

    o   The Math group generic does not check the number of arguments
    supplied before dispatch: it used to if the default method had
    one argument but not if it had two.  This allows trunc.POSIXt()
    to be called via the group generic trunc().

    o   Logical matrix replacement indexing of data frames is now
    implemented (interpreted as if the lhs was a matrix).

    o   Recursive indexing of lists is allowed, so x[[c(4,2)]] is
    shorthand for x[[4]][[2]] etc.  (Wishlist PR#1588)

    o   Most of the time series functions now check explicitly for a
    numeric time series, rather than fail at a later stage.

    o   The postscript output makes use of relative moves, and so is
    somewhat more compact.

    o   %*% and crossprod() for complex arguments make use of BLAS
    routines and so may be much faster on some platforms.

    o   arima() has coef(), logLik() (and hence AIC) and vcov() methods.

    o   New function as.difftime() for time-interval data.

    o   basename() and dirname() are now vectorized.

    o   biplot.default() {mva} allows `xlab' and `ylab' parameters to
    be set (without partially matching to `xlabs' and `ylabs').
    (Thanks to Uwe Ligges.)

    o   New function capture.output() to send printed output from an expression
    to a connection or a text string.

    o   ccf() (pckage ts) now coerces its x and y arguments to class "ts".

    o   chol() and chol2inv() now use LAPACK routines by default.

    o   as.dist(.) is now idempotent, i.e., works for "dist" objects.

    o   Generic function confint() and `lm' method (formerly in
    package MASS, which has `glm' and `nls' methods).

    o   New function constrOptim() for optimisation under linear inequality
    constraints.

    o   Add `difftime' subscript method and methods for the group
    generics. (Thereby fixing PR#2345)

    o   download.file() can now use HTTP proxies which require `basic'
    username/password authentication.

    o   dump() has a new argument `envir'.  The search for named
    objects now starts by default in the environment from which
    dump() is called.

    o   The edit.matrix() and edit.data.frame() editors can now handle
    logical data.

    o   New argument `local' for example() (suggested by Andy Liaw).

    o   New function file.symlink() to create symbolic file links
    where supported by the OS.

    o   New generic function flush() with a method to flush connections.

    o   New function force() to force evaluation of a formal argument.

    o   New functions getFromNamespace(), fixInNamespace() and
    getS3method() to facilitate developing code in packages with
    namespaces.

    o   glm() now accepts `etastart' and `mustart' as alternative ways
    to express starting values.

    o   New function gzcon() which wraps a connection and provides
    (de)compression compatible with gzip.

    load() now uses gzcon(), so can read compressed saves from
    suitable connections.

    o   help.search() can now reliably match individual aliases and
    keywords, provided that all packages searched were installed
    using R 1.7.0 or newer.

    o   hist.default() now returns the nominal break points, not those
    adjusted for numerical tolerances.

    To guard against unthinking use, `include.lowest' in
    hist.default() is now ignored, with a warning, unless `breaks'
    is a vector.  (It either generated an error or had no effect,
    depending how prettification of the range operated.)

    o   New generic functions influence(), hatvalues() and dfbeta()
    with lm and glm methods; the previously normal functions rstudent(),
    rstandard(), cooks.distance() and dfbetas() became generic.
    These have changed behavior for glm objects -- all originating from
    John Fox' car package.

    o   interaction.plot() has several new arguments, and the legend
    is not clipped anymore by default.  It internally uses axis(1,*)
    instead of mtext().
    This also addresses "bugs" PR#820, PR#1305, PR#1899.

    o   New isoreg() function and class for isotonic regression
    (`modreg' package).

    o   La.chol() and La.chol2inv() now give interpretable error
    messages rather than LAPACK error codes.

    o   legend() has a new `plot' argument.  Setting it `FALSE' gives
    size information without plotting (suggested by U.Ligges).

    o   library() was changed so that when the methods package is
    attached it no longer complains about formal generic functions
    not specific to the library.

    o   list.files()/dir() have a new argument `recursive'.

    o   lm.influence() has a new `do.coef' argument allowing *not* to
    compute casewise changed coefficients.  This makes plot.lm() much
    quicker for large data sets.

    o   load() now returns invisibly a character vector of the names
    of the objects which were restored.

    o   New convenience function loadURL() to allow loading data files
    from URLs (requested by Frank Harrell).

    o   New function mapply(), a multivariate lapply().

    o   New function md5sum() in package tools to calculate MD5
    checksums on files (e.g. on parts of the R installation).

    o   medpolish() {package eda} now has an `na.rm' argument (PR#2298).

    o   methods() now looks for registered methods in namespaces, and
    knows about many objects that look like methods but are not.

    o   mosaicplot() has a new default for `main', and supports the
    `las' argument (contributed by Uwe Ligges and Wolfram Fischer).

    o   An attempt to open() an already open connection will be detected
    and ignored with a warning.  This avoids improperly closing
    some types of connections if they are opened repeatedly.

    o   optim(method = "SANN") can now cover combinatorial optimization
    by supplying a move function as the `gr' argument (contributed
    by Adrian Trapletti).

    o   PDF files produced by pdf() have more extensive information
    fields, including the version of R that produced them.

    o   On Unix(-alike) systems the default PDF viewer is now determined
    during configuration, and available as the 'pdfviewer' option.

    o   pie(...) has always accepted graphical pars but only passed
    them on to title().  Now pie(, cex=1.5) works.

    o   plot.dendrogram (`mva' package) now draws leaf labels if present
    by default.

    o   New plot.design() function as in S.

    o   The postscript() and PDF() drivers now allow the title to be set.

    o   New function power.anova.test(), contributed by Claus Ekstrøm.

    o   power.t.test() now behaves correctly for negative delta in the
    two-tailed case.

    o   power.t.test() and power.prop.test() now have a `strict'
    argument that includes rejections in the "wrong tail" in the
    power calculation. (Based in part on code suggested by Ulrich
    Halekoh.)

    o   prcomp() is now fast for n x m inputs with m >> n.

    o   princomp() no longer allows the use of more variables than
    units: use prcomp() instead.

    o   princomp.formula() now has principal argument `formula', so
    update() can be used.

    o   Printing an object with attributes now dispatches on the
    class(es) of the attributes. See ?print.default for the fine
    print. (PR#2506)

    o   print.matrix() and prmatrix() are now separate functions.
    prmatrix() is the old S-compatible function, and
    print.matrix() is a proper print method, currently identical
    to print.default().  prmatrix() and the old print.matrix()
    did not print attributes of a matrix, but the new print.matrix()
    does.

    o   print.summary.{lm,glm} now default to symbolic.cor = FALSE, but
    symbolic.cor can be passed to the print methods from the
    summary methods.  print.summary.{lm,glm} print correlations to
    2 decimal places, and the symbolic printout avoids abbreviating
    labels.

    o   If a prompt() method is called with 'filename' as 'NA', a
    list-style representation of the documentation shell generated
    is returned.  New function promptData() for documenting objects
    as data sets.

    o   qqnorm() and qqline() have an optional logical argument
    `datax' to transpose the plot (S-PLUS compatibility).

    o   qr() now has the option to use LAPACK routines, and the
    results can be used by the helper routines qr.coef(), qr.qy()
    and qr.qty().  The LAPACK-using versions may be much faster
    for large matrices (using an optimized BLAS) but are less
    flexible.

    o   QR objects now have class "qr", and solve.qr() is now just the
    method for solve() for the class.

    o   New function r2dtable() for generating random samples of two-way
    tables with given marginals using Patefield's algorithm.

    o   rchisq() now has a non-centrality parameter `ncp', and there's a
    C API for rnchisq().

    o   New generic function reorder() with a dendrogram method;
    new order.dendrogram() and heatmap().

    o   require() has a new argument, character.only,
      -- to make it align with library.

    o   New functions rmultinom() and dmultinom(), the first one with
    a C API.

    o   New function runmed() for fast runnning medians (`modreg' package).

    o   New function slice.index() for identifying indexes with respect
    to slices of an array.

    o   solve.default(a) now gives the dimnames one would expect.

    o   stepfun() has a new `right' argument for right-continuous step
    function construction.

    o   str() now shows ordered factors different from unordered ones.
    It also differentiates "NA" and as.character(NA), also for factor
    levels.

    o   symnum() has a new logical argument `abbr.colnames'.

    o   summary(<logical>) now mentions NA's as suggested by
    Göran Broström.

    o   summaryRprof() now prints times with a precision appropriate
    to the sampling interval, rather than always to 2dp.

    o   New function Sys.getpid() to get the process ID of the R session.

    o   table() now allows exclude= with factor arguments (requested by
    Michael Friendly).

    o   The tempfile() function now takes an optional second argument
    giving the directory name.

    o   The ordering of terms for terms.formula(keep.order=FALSE) is
    now defined on the help page and used consistently, so that
    repeated calls will not alter the ordering (which is why
    delete.response() was failing: see the bug fixes).  The
    formula is not simplified unless the new argument `simplify'
    is true.

    o   added "[" method for terms objects.

    o   New argument `silent' to try().

    o   ts() now allows arbitrary values for y in start/end = c(x, y):
    it always allowed y < 1 but objected to y > frequency.

    o   unique.default() now works for POSIXct objects, and hence so
    does factor().


    o   Package tcltk now allows return values from the R side to the
    Tcl side in callbacks and the R_eval command. If the return
    value from the R function or expression is of class "tclObj"
    then it will be returned to Tcl.

    o   A new HIGHLY EXPERIMENTAL graphical user interface using the tcltk
    package is provided. Currently, little more than a proof of concept.
    It can be started by calling "R -g Tk" (this may change in later
    versions) or by evaluating tkStartGUI(). Only Unix-like systems
    for now. It is not too stable at this point; in particular, signal
    handling is not working properly.


    o   Changes to support name spaces:

    - Placing base in a name space can no longer be disabled by
      defining the environment variable R_NO_BASE_NAMESPACE.

    - New function topenv() to determine the nearest top level
      environment (usually .GlobalEnv or a name space environment).

    - Added name space support for packages that do not use methods.

    o   Formal classes and methods can be `sealed', by using the
    corresponding argument to setClass or setMethod.  New
    functions isSealedClass() and isSealedMethod() test sealing.

    o   packages can now be loaded with version numbers. This allows
    for multiple versions of files to be installed (and potentially
    loaded). Some serious testing will be going on, but it should
    have no effect unless specifically asked for.


INSTALLATION CHANGES

    o   TITLE files in packages are no longer used, the Title field
    in the DESCRIPTION file being preferred.  TITLE files will be
    ignored in both installed packages and source packages.

    o   When searching for a Fortran 77 compiler, configure by default
    now also looks for Fujitsu's frt and Compaq's fort, but no
    longer for cf77 and cft77.

    o   Configure checks that mixed C/Fortran code can be run before
    checking compatibility on ints and doubles: the latter test
    was sometimes failing because the Fortran libraries were not
    found.

    o   PCRE and bzip2 are built from versions in the R sources if the
    appropriate library is not found.

    o   New configure option --with-lapack to allow high-performance
    LAPACK libraries to be used: a generic LAPACK library will be
    used if found.  This option is not the default.

    o   New configure options --with-libpng, --with-jpeglib, --with-zlib,
    --with-bzlib and --with-pcre, principally to allow these
    libraries to be avoided if they are unsuitable.

    o   If the precious variable R_BROWSER is set at configure time
    it overrides the automatic selection of the default browser.
    It should be set to the full path unless the browser appears
    at different locations on different client machines.

    o   Perl requirements are down again to 5.004 or newer.

    o   Autoconf 2.57 or later is required to build the configure
    script.

    o   Configure provides a more comprehensive summary of its results.

    o   Index generation now happens when installing source packages
    using R code in package tools.  An existing 'INDEX' file is used
    as is; otherwise, it is automatically generated from the \name
    and \title entries in the Rd files.  Data, demo and vignette
    indices are computed from all available files of the respective
    kind, and the corresponding index information (in the Rd files,
    the 'demo/00Index' file, and the \VignetteIndexEntry{} entries,
    respectively).  These index files, as well as the package Rd
    contents data base, are serialized as R objects in the 'Meta'
    subdirectory of the top-level package directory, allowing for
    faster and more reliable index-based computations (e.g., in
    help.search()). For vignettes an HTML index is generated
    and linked into the HTML help system.

    o   The Rd contents data base is now computed when installing source
    packages using R code in package tools.  The information is
    represented as a data frame without collapsing the aliases and
    keywords, and serialized as an R object.  (The 'CONTENTS' file
    in Debian Control Format is still written, as it is used by the
    HTML search engine.)

    o   A NAMESPACE file in root directory of a source package is copied
    to the root of the package installation directory.  Attempting to
    install a package with a NAMESPACE file using --save signals an
    error; this is a temporary measure.

    o   The defaults for configure for Darwin systems is
    --with-blas='-framework vecLib' --with-lapack --with-aqua
    that by default builds R as a framework and installs it in
    /Library/Frameworks as R.framework. Then, make install just
    installs the R.framework in /Library/Frameworks unless specified
    at configure time using the -enable-R-framework=[DIR] or using
    the --prefix flag at installation time.

DEPRECATED & DEFUNCT

    o   The assignment operator `_' will be removed in the next
    release and users are now warned on every usage: you may even see
    multiple warnings for each usage.

    If environment variable R_NO_UNDERLINE is set to anything of
    positive length then use of `_' becomes a syntax error.

    o   machine(), Machine() and Platform() are defunct.

    o   restart() is defunct.  Use try(), as has long been recommended.

    o   The deprecated arguments `pkg' and `lib' of system.file() have
    been removed.

    o   printNoClass() {methods} is deprecated (and moved to base,
    since it was a copy of a base function).

    o   Primitives dataClass() and objWithClass() have been replaced
    by class() and class<-(); they were internal support functions
    for use by package methods.

    o   The use of SIGUSR2 to quit a running R process under Unix is
    deprecated, the signal may need to be reclaimed for other
    purposes.


UTILITIES

    o   R CMD check more compactly displays the tests of DESCRIPTION
    meta-information.  It now reports demos and vignettes without
    available index information.  Unless installation tests are
    skipped, checking is aborted if the package dependencies cannot
    be resolved at run time.  Rd files are now also explicitly
    checked for empty \name and \title entries.  The examples are
    always run with T and F redefined to give an error if used
    instead of TRUE and FALSE.

    o   The Perl code to build help now removes an existing example
    file if there are no examples in the current help file.

    o   R CMD Rdindex is now deprecated in favor of function Rdindex()
    in package tools.

    o   Sweave() now encloses the Sinput and Soutput environments of
    each chunk in an Schunk environment. This allows to fix some
    vertical spacing problems when using the latex class slides.


C-LEVEL FACILITIES

    o   A full double-precision LAPACK shared library is made
    available as -lRlapack.  To use this include
    $(LAPACK_LIBS) $(BLAS_LIBS) in PKG_LIBS.

    o   Header file R_ext/Lapack.h added.  C declarations of BLAS
    routines moved to R_ext/BLAS.h and included in R_ext/Applic.h
    and R_ext/Linpack.h for backward compatibility.

    o   R will automatically call initialization and unload routines, if
    present, in shared libraries/DLLs during dyn.load() and
    dyn.unload() calls. The routines are named R_init_<dll name>
    and R_unload_<dll name>, respectively.  See the Writing R
    Extensions Manual for more information.

    o   Routines exported directly from the R executable for use with
    .C(), .Call(), .Fortran() and .External() are now accessed via
    the registration mechanism (optionally) used by packages.  The
    ROUTINES file (in src/appl/) and associated scripts to
    generate FFTab.h and FFDecl.h are no longer used.

    o   Entry point Rf_append is no longer in the installed headers
    (but is still available).  It is apparently unused.

    o   Many conflicts between other headers and R's can be avoided by
    defining STRICT_R_HEADERS and/or R_NO_REMAP -- see `Writing R
    Extensions' for details.

    o   New entry point R_GetX11Image and formerly undocumented
    ptr_R_GetX11Image are in new header R_ext/GetX11Image.  These
    are used by package tkrplot.


BUG FIXES

    o   The redefinition of the internal do_dataentry by both the aqua and X11
    modules, casued a bus error when launching R without the --gui=aqua option
    under X11 using a version of R built to use the aqua module. This has
    now been fixed. (PR#6438)

    o   Sys.sleep() on Unix was having trouble with waits of less than 0.5s

    o   The fix to PR#2396 broke read.table() on files with CR line
    endings. (PR#2469)   Separate problem with this on Carbon Mac OS
    build fixed as well.

    o   Converting Sweave files to noweb syntax using SweaveSyntConv()
    was broken.

    o   Printing numbers near the minimum could get the number of
    significant figures wrong due to underflow: for example 4e-308
    might print as 4.00000e-308.  (Seen on some Windows builds,
    and also on numbers around 1e-317 on Linux.)

    o   wilcox.test() could give integer overflow warnings on very long
    vectors.  Also added tests for numeric inputs, as per the help
    page.  (PR#2453)

    o   Printing unquoted character vectors containing escape
    characters was computing the wrong length and hence
    misaligning names.  This was due to a bug in Rstrlen which
    might have had other effects.

    o   if(logical(0)) and while(logical(0)) now report zero length,
    not `missing value where logical is needed'.

    o   The gaussian() and inverse.gaussian() families were documented
    to allow only one link, which has not been true in R for at
    least four years.

    o   prmatrix() forced conversion to character if `na.print' was
    used, and that conversion neither respected `digits' nor
    `quote'.

    o   Rprof() might give misleading results for too small values of
    `interval' and in practice the default 20ms was about as small
    as is advisable on Linux.  Now the interval is forced to be at
    least one clock tick.

    o   summary.data.frame() was not giving interpretable results when
    the data frame contained a data frame as a column. (PR#1891)

    o   delete.response() might re-order the rhs terms so prediction
    might fail or even give incorrect results. (PR#2206)

    o   StructTS() now accepts numeric time series of integer storage mode.

    o   all(), any() now handle NAs as documented.

    o   Subsetting arrays to a result with 0 dimension(s) failed if
    the array had dimnames. (PR#2507)

    o   If the call to data.frame() included 0-row arguments, it tried
    to replicate them to the maximum number of rows, and failed if
    this was 1 or more.

    o   replicate() now understands data frames to which na.omit() has
    been applied.

    o   is.ts() was too liberal: a time series must have at least one point.

    o   methods() was sorting by package, not by name.

    o   symbols(thermometers=) was often giving a spurious warning about
    the range.

    o   tcltk was using deprecated internals of the Tcl library when
    accessing error messages.  Not likely to be a user-visible
    change.

    o   The automatic search for BLAS libs now tries Sun's libsunperf
    the way the latest versions require. (PR#2530)

    o   str(array(1)) now does show the array.
    str(Surv(...)) now works again.

    o   step(), add1.default() and drop1.default() now work somewhat
    better if called from a function.

    o   page() was searching from the wrong environment, and so
    searching base before the workspace.

    o   crossprod(Z) for complex Z was returning nonsense.

    o   La.chol2inv() gave incorrect results unless the matrix was square.

    o   When the POSIXt date functions were required to guess DST,
    they sometimes guessed correctly that DST was in force but
    converted a POSIXlt time as if standard time was given.

    o   c/rbind were not handling zero col/row matrices correctly.
    (PR#2541 was one symptom.)

    o   approx() and approxfun() now work with 1 knot if
    method = "constant".  stepfun(), ecdf() and plot.stepfun() do so
    as well.

    o   AIC.lm/default was failing if multiple objects and k were
    specified.  (PR#2518)

    o   removeMethods{methods} was broken. (PR#2519)

    o   summary.glm() had two `aic' components in the returned object.

    o   autoload() was returning the value of its last command, a
    promise, even though it was documented to have no value.
    As a result some packages (e.g. nlme) were loading packages
    they meant to autoload.

    o   Fixes to methods and classes:
       - show() is consistent with using setOldClass for S3 classes.
       - several problems with the coerce and replace methods
         generated by setIs have been fixed.
       - more thorough tests & informative messages for invalid
         `def' arguments to setGeneric
       - setGeneric will now create the generic function even when
         a generic of the same name already exists (it does issue
         a warning).

    o   unz() connections could no longer be opened. (PR#2579)

    o   unique(ordered factor) returned an unordered factor. (PR#2591)

    o   x[] <- value coerced x to the mode of value if and only if x
    had length 0!  (Should only happen if x is null: PR#2590)

    o   lm() mislabelled the cols of the qr decomposition. (cause of PR#2586)

    o   data() looks for file extensions in an order prescribed in the
    help file: previously whether foo.R or foo.csv was used was
    locale-dependent.

    o   sys.function() now returns the actual function being evaluated in
    the specified frame rather than one inferred from the call.

    o   match.call() now uses the definition of the actual function being
    evaluated rather than one inferred from the call.

    o   abbreviate(*, dot = TRUE) now only adds a "." where abbreviations
    did happen.

    o   Changing timezones in the POSIXt functions was not working on
    some Linux systems, and this has been corrected.

    o   ks.test() in package ctest had numerical problems in the lower
    tail of the asymptotic distribution (PR#2571).

    o   Sweave() now handles empty chunks at the end of files correctly.

    o   [<-() lost the object bit if coercion was involved.

    o   package::object wasn't being deparsed properly.

    o   seq.POSIXt() with `by' an object of class "difftime" ignored
    the units.

    o   rank(c("B", NA)) no longer returns character.

    o   reference to by() added in ?tapply

    o   ?lm describes what happens with matrix response

    o   The X11 device has improved event handling. In particular it
    used to often miss the last of a series of resize events.

    o   lm.influence() and related functions now work again for the
    multivariate case and when there are zero weights.

    o   format( <character> ) now always keeps names and dimnames.

    o   table(factor(c(2,NA), exclude=NULL)) prints better now.

    o   predict(foo, type = "terms") and hence
    residuals(foo, type = "partial") now work for lm and glm objects
    with weights zero.  Further, model.matrix() is now only called once.

    o   R CMD config now works correctly when called from a Makefile
    using GNU make.

    o   The data.frame method for rbind() was
    - converting character columns to factors,
    - converting ordered factor columns to unordered factors,
    - failing to append correctly a factor to a character column
      and vice versa.

    o   as.hclust.twins() now does provide proper `labels', `method' and
    `call' components.

    o   cycle() sometimes failed on a time series which started at a cycle
    other than 1.

    o   read.dcf() read incorrectly files which did not end in a new line.

    o   read.socket() dropped certain non-alphanumeric characters. (PR#2639)

    o   termplot() handles missing data better (PR#2687,
     <Mark.Bravington@csiro.au>)

    o   Corrected MacRoman encoding for Icircumflex etc.



    **************************************************
    *                        *
    *          1.6 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 1.6.2


BUG FIXES

    o   plot.stepfun() now obeys a `ylim=.' specification.

    o   removeClass() does a better job of removing inheritance
    information.

    o   setIs() will not allow mismatched representations between two
    classes (without an explicit coerce method).

    o   The code underlying polygon drawing contained a memory leak.
    This showed up in persp, but did not affect other graphics
    functions.  It is now possible to draw big DEMs.

    o   logLik.nls() gave wrong df. (PR#2295)

    o   rbind() with a mixture of data frames and matrices treated the
    matrices as vectors. (PR#2266)

    o   stripchart(method="stack") was not handling missing values. (PR#2018)

    o   Arithmetic functions such as log() lost the object bit from
    classed objects if coercion was needed. (PR#2315)

    o   exp_rand would go into an infinite loop if unif_rand returned 0.

    o   formatC(x, format="fg") could return exponential format if
    rounding pushed x over a positive power of 10. (PR#2299)

    o   attr(x, foo) used partial matching for `foo' (even though not
    documented to do so), and failed to find `foo' if there were
    two or more partial matches before the exact match in the list
    of attributes.

    o   Rdconv now creates direct HTML hyperlinks when linking to
    documentation in the same package.  The code now ensures that
    links which can be resolved within the package are so resolved,
    even when there are possible resolutions in other packages.

    o   If readBin(what=character()) is used incorrectly on a file which
    does not contain C-style character strings, warnings (usually
    many) are now given.

    o   Building libR.so with the zlib in the R sources was not
    finding the local zlib headers.

    o   system(intern=TRUE) has an undocumented line length limit of
    119 chars both on Unix and Windows.  The limit is now 8096 and
    documented.  On Unix (only) every 120th character used to be
    discarded.

    o   plot.POSIX[cl]t were not passing graphics parameters on to
    axis.POSIXct.

    o   On some HP-UX systems, installed scripts were not executable
    when using the BSD-compatible install system program found by
    configure.  We now always use install-sh on HP-UX. (PR#2091)

    o   c() was converting NA names to "NA": now proper NA strings are
    used wherever possible. (PR#2358)

    o   Checks in the C code prevent possible memory faults when
    standardGeneric is called invalidly.

    o   Macros NEW_OBJECT (aka NEW) and MAKE_CLASS added; required by
    the .Call interface to generate arbitrary objects.

    o   A typo was causing segfaults when using data.entry under SuSE.

    o   mostattributes<-() was failing to copy across dimnames when
    one component was NULL, affecting pmax() and pmin() when the
    first argument was a matrix.  (root cause of PR#2357)

    o   The pdf() device now initialises graphical parameters
    properly.  (PR#2281)

    o   Checks in the C code prevent possible memory faults when
    standardGeneric is called invalidly.

    o   Macros NEW_OBJECT (aka NEW) and MAKE_CLASS added; required by
    the .Call interface to generate arbitrary objects.

    o   Problem that prevented package tcltk from working with Tcl/Tk
    8.4 (crash on initialization) resolved.  (Notice that binaries
    may still require an older Tcl/Tk, for example on Windows).

    o   type.convert() was not getting the levels right if passed a
    character vector containing <NA>s, and `na.strings' did not
    contain "NA".  This affected read.table().

    o   Internal match function did not check for nor handle 0-length
    vectors.  (The R function match() did.)  This could cause
    type.convert() to segfault.

    o   The line length limit in output text connections has been
    raised to 8095 chars.

    o   Sweave now uses anonymous file rather than text connections
    to avoid the limits of the latter (see previous item).

    o   parsing did not work on connections when pushback was used (as
    it had never been implemented).  (PR#2396)

    o   max.col() only found NAs in the first column (typo).

    o   Added a workaround for recent versions of glibc (e.g. RedHat 8.0)
    with inconsistent mktime/localtime functions which caused
    conversion to/from POSIXct times prior to 1970-01-01 to be
    inconsistent.  On such platforms this is a run-time test to
    allow e.g. R compiled on RH7.2 to run on RH8.0.

    o   Clipping was not being reset properly between plots on the gtk()
    device (the default under the GNOME interface). (PR#2366)

    o   axis(*, fg= cc) now works (again) the same as axis(*, col = cc).



        CHANGES IN R VERSION 1.6.1


NEW FEATURES

    o   Added a few "trivial and obviously missing" functions to tcltk:
    tkchooseDirectory, tkpopup, tkdialog, tkread

    o   barplot() has a new argument `axis.lty', which if set to 1
    allows the pre-1.6.0 behaviour of plotting the axis and tick
    marks for the categorical axis.  (This was apparently not
    intentional, but axis() used to ignore lty=0.)
    The argument `border' is no longer ".NotYetUsed".


BUG FIXES

    o   hist(<datetime>, cex.axis = f)  now works for x-axis too.

    o   prompt() gave wrong \usage{.} for long argument default expressions.

    o   summary(x) gives more information when `x' is a logical
    (or a data frame with a logical column which is now quite customary).

    o   seq.POSIXt(from, to, length.out= . ) could give too long results

    o   summaryRprof() was counting nested calls to the same function
    twice.

    o   Printing of objects of mode "expression" did strange things if
    there were "%" characters in the deparsed expression (PR#2120).

    o   as.matrix.data.frame converted missings to "NA" not character NA.
    (PR#2130)

    o   spec.pgram() was only interpolating zero freq for one series. (PR#2123)

    o   help(randu) had % unescaped in the example. (PR#2141)

    o   Making html links would fail if packages-head.html was not
    writeable. (PR#2133)

    o   Sweave.sty was not installed to $R_HOME/share/texmf when
    builddir != srcdir. On Windows backslashes in latex paths have
    to be replaced by slashes.

    o   A memory leak in deparsing was introduced when eliminating
    static variables (thanks to Achim Zeileis for spotting this).
    A similar problem in loading workspaces has been corrected.

    o   TclInterface.Rd incorrectly used \synopsis for \usage so that
    the usage section wasn't output.

    o   Readline stack off-by-one error. (PR#2165)

    o   R_ExpandFileName had a memory leak in the case libreadline was
    used under Unix-alikes.

    o   sys.save.image() now closes all connections so it will work even
    if the connection list has become full.

    o   loess() had an unstated limit of four predictors: this is now
    documented and enforced.

    o   ${R_HOME}/etc/Renviron.site is now not read if R_ENVIRON is
    set, as documented.  Previously it was read unless R_ENVIRON
    pointed to an actual file.

    o   Startup.Rd described the processing under Unix-alikes but
    incorrectly implied it happened that way on the Windows and
    Mac OS ports.  Neither use Renviron.site, for example.

    o   besselK(x,*) now returns 0 instead of Inf for large x. (PR#2179)

    o   The Tcl console code didn't work with Tcl/TK 8.0, and has been
    #ifdef'd out.  (PR#2090)

    o   format.AsIs() was not handling matrices.

    o   sd() was not passing na.rm to var() for matrices and data frames.

    o   dist() {mva} silently treated +/-Inf as NA.

    o   setwd() now returns NULL invisibly.

    o   basename() and dirname() did not check the length of their
    input and ignored elements after the first.  This affected
    undoc {tools}.

    o   If A had dimnames, eigen(A) had inappropriate dimnames. (PR#2116)

    o   as.POSIXct.dates had a sign error for the origin (PR#2222)

    o   The claim that pie charts should be avoided (in pie.Rd) is now
    supported by a quote from Cleveland (1985).

    o   The vsnprintf() functions supplied for systems that don't
    supply their own had a bug in the output of fractional parts,
    corrupting data if using save() with ascii=TRUE. (PR#2144)

    o   pretty() values close to 0 in some cases which are now 0 (PR#1032
    and D.Brahm's mails).

    o   factor.scope() was giving an error rather than a reasonable
    default in some cases when it was called incorrectly (e.g. via
    incorrect formula in add1()).


BUILD ISSUES

    o   Toplevel Makefile was missing dependency of "docs" on "R" (causing
    parallel makes to go wrong)

    o   When building with recommended packages those were installed
    into the first path in R_LIBS, if the environment variable was
    present.



        CHANGES IN R VERSION 1.6.0


USER-VISIBLE CHANGES

    o   The default colour palette now has "grey" instead of "white"
    in location 8.  See palette().

    o   grid(nx) behaves differently (but the same as in R versions <= 0.64).


NEW FEATURES

    o   New operator :: in the grammar, for name spaces.

    o   New faster rowsum(), also works on data frames.

    o   grep(), (g)sub() and regexpr() have a new argument `perl'
    which if TRUE uses Perl-style regexps from PCRE (if installed).
    New capabilities option "PCRE" to say if PCRE is available.

    o   Preparations for name space support:

    - Functions in the base package are now defined in a name space.
      As a temporary measure, you can disable this by defining the
      environment variable R_NO_BASE_NAMESPACE.

    - UseMethod dispatching now searches for methods in the
      environment of the caller of the generic function rather
      than the environment where the generic is defined.

    o   The objects created in the methods package to represent
    classes, generic functions, method definitions, and
    inheritance relations now themselves belong to true classes.  In
    particular, the "classRepresentation" objects follow the description
    in "Programming with Data" (section 7.6).

    o   Other additions and changes to the methods package:

    - The function setOldClass() has been added, following the
      description on page 450 of "Programming with Data".  Use it
      if old-style classes are to be supplied in signatures for
      setMethod, particularly if the old-style classes have
      inheritance.  Many of the old-style classes in the base
      package should be pre-specified; try getClass("mlm"), e.g.

    - The setGeneric() function applies some heuristics to warn
      about possibly erroneous generic function definitions.
      (Before, obscure bugs could result.)

    - The function promptMethods() has been revised to work better
      and to provide aliases for individual methods.

    - The behavior of the as() function has been generalized, in
      particular with a strict= argument, the general goal being
      to let simple extensions of classes pass through in method
      dispatch and related computations without altering the
      objects. More to make method behavior more "natural" than
      for direct use.

    - Some inconsistencies following detach("package:methods")
      have been removed, so it _should_ be possible to
      detach/re-attach the methods package.

    o   New methods ([[, print, str) and extended plot() method
    (incl. logical `horiz') for "dendrogram" class.

    o   sprintf() now checks the agreement between formats and object
    types, and handles special values (NA, Inf, ...) correctly.

    o   chol() now uses a tolerance for non-positive-definiteness and
    so should give more consistent results across platforms.

    o   New function agrep() for approximate (fuzzy) string matching.

    o   help.search() can now use both approximate (fuzzy) and regular
    expression matching.  By default, if the pattern to be matched
    consists of only alphanumeric characters, whitespace or a dash,
    approximate matching is used.

    o   axis() has three new optional arguments `col', `lty', and `lwd'
    all for drawing the axis line and tick marks.

    o   Function vcov() (formerly in MASS), a generic function to
    return the variance-covariance matrix of the parameter
    estimates of a fitted model.

    o   duplicated() and unique() have methods for matrices and arrays
    (based on ideas from Jens Oehlschlägel).

    o   Internally memory sizes and counts of cons cells are now stored
    in unsigned longs.  This allows memory limits to be set and
    objects created in the range 2-4Gb on 32-bit platforms, and
    allows 64-bit platforms to use much larger amounts of memory.

    o   Command-line flags to set memory can now use the suffix `G'
    for gigabytes.  The setting of maximum vsize is now only
    limited by the platform's address space.

    o   All warning and error messages are truncated to a length set
    by options(warning.length=), defaulting to 1000.  (Previously
    most (but not quite all) were truncated at 8192 characters.)

    o   [dpqr]gamma() check for shape parameter > 0.

    o   as.POSIX[cl]t can now convert logical NAs.

    o   All installed packages (even those shipped with R) are
    given a `Built' field in the DESCRIPTION file.

    o   as.data.frame() now coerces logical matrices into logical
    columns (rather than factors).

    o   [[<-.data.frame no longer coerces character replacement values
    to factor.  This is consistent with using $ to replace and
    with S4.

    o   library() attempts to detect improperly installed packages, so
    as from this version an installed package must have a
    DESCRIPTION file and that file must have been stamped with a
    `Built:' line (which was introduced in 1.2.0).  Under
    Unix-alikes, the platform is checked against that used for
    installation.

    o   print.factor() has new arguments `max.levels' (with a smart default)
    and `width'.  print.ordered() is no longer needed.

    o   RNGkind() has an additional option for normal random generators:
    "Inversion".

    o   data.frame() recycles factors and "AsIs" objects as well as
    atomic vectors.

    o   predict.lm() warns if `newdata' is supplied and the fit was
    rank-deficient, as this can be misleading.

    o   rect() accepts additional graphics parameters through a ...
    argument (in the same way as polygon).

    o   strwidth/strheight() now coerce their first argument in exactly
    the same way text() does, so a wider range of inputs is allowed.

    o   prompt()'s default and data.frame methods have a new 3rd argument
    `name' allowing them to used more easily in scripts and loops.

    o   rgb() has a new `maxColorValue' argument, allowing r,g,b in [0,M],
    particularly in {0:255}, efficiently and non-error-prone.

    o   summaryRprof() provides the functionality of R CMD Rprof in R
    code, though more slowly.

    o   stop() accepts multiple arguments (which are concatenated)
    just as warning() does.

    o   scan() now throws an error with incorrect logical input (which
    was previously taken as FALSE).

    o   pdf() now uses PDF not R code for clipping, which ensures that
    partially visible text strings are (partially) shown.

    o   Each R session uses a per-session temporary directory which
    is removed at normal termination.  The directory name is given
    by the tempdir() function, and filenames returned by
    tempfile() will be within that directory.

    o   help.start() on Unix now uses a .R subdirectory of the
    per-session temporary directory and not ~/.R.  A side effect
    is that ~/.R is now never deleted by R.

    This now uses the remote control mechanism only if the X
    display is local to the R process (as otherwise it might use a
    browser running on an arbitrary machine).

    o   *Very* experimental browseEnv() for browsing objects in an
    environment.

    o   cbind/rbind() used to ignore all zero-length vectors, an
    undocumented quirk for S-compatibility.  This caused problems
    when combining zero-extent matrices and zero-length vectors, and
    now zero-length vectors are ignored unless the result would
    have zero rows/columns.

    o   read.table(stdin()) will now work.

    o   plot.spec(x) now also works for other x than AR and Pgram results.

    o   New functions La.chol() and La.chol2inv() for Cholesky
    decomposition and inverse of positive definite matrices using
    Lapack.

    o   Changes to the tcltk package

    - on Unix systems, the Tcl event loop has been integrated with
      R's own (so that tkwait.variable() no longer halts updates of
      plot windows).

    - also on Unix, stubs have been created to divert R's input
      and output routines to go via Tcl commands. (Nothing
      uses this at present, but packages might be developed to
      take advantage of it.)

    - return value from Tcl commands is no longer invisible. A new
      print method, print.tclObj(), has been introduced.

    - Tcl variables created by tclVar() are now explicitly put into
      Tcl's global namespace, removing potential scoping problems.

    - The tcltk dynamic library now loads with local=FALSE since
      the default had trouble when loading Tcl extensions
      (e.g. Tix)

    - The tkpager() function had not been updated for the return
      value change from 1.5.0

    o   The bmp(), jpeg() and png() devices can produce multiple
    bitmap files, one for each page.  The default filenames have been
    changed to include a sequence number.

    o   New function axTicks() returning tick mark locations like axis().

    o   grid() has a more sensible default behavior.  Tick axis alignment
    only happens when no numbers of grid cells are specified.  New
    arguments lwd and equilogs; nx/ny = NA for not drawing, see ?grid.

    o   installed.packages() has a new argument `priority'.

    o   termplot() uses factor levels rather than 1,2,3... for x-axis.

    o   Workaround for optimization bugs on gcc 3.1/2 on 32-bit Solaris.

    o   The trace() function has been robustified and a new function
    tracingState() added to turn tracing temporarily on and off.

    o   New cophenetic() in "mva" as utility for hierarchical clustering.

    o   p.adjust() has two new methods, 'Hommel' and 'FDR', contributed
    by Gordon Smyth <smyth@wehi.edu.au>.

    o   stars() now has add and plot arguments.


DEPRECATED & DEFUNCT

    o   The assignment operator `_' is deprecated: a warning is given
    once per R session.

    o   machine() is deprecated in favour of .Platform$OS.type.
    Machine() and Platform() are deprecated in favour of .Machine
    and .Platform.

    o   arima0.diag() (package ts) is defunct.

    o   piechart() is defunct.

    o   print.ordered() has been removed, so print.factor() is used.

    o   The global internal variables .Dyn.libs and .lib.loc are
    removed in favor of the internal functions .dynLibs() and
    .libPaths().

    o   restart() is deprecated in preparation for proper exception
    handling.  Use try(), as has long been recommended.


DOCUMENTATION CHANGES

    o   New demo(persp) containing some of the former example(persp) ones
    and more.


C-LEVEL FACILITIES

    o   Rversion.h is no longer automatically included by R.h.
    Include it explicitly if you need it.

    o   New entry point R_tmpnam in Utils.h.

    o   The Unix event loop interface has been changed to facilitate
    integration with other loops. R_checkActivity and
    R_runHandlers should eventually replace getSelectedHandler.


INSTALLATION CHANGES

    o   Perl 5.005 or newer is now required.

    o   R CMD INSTALL is now guaranteed to sort the R source files in
    ASCII order.


UTILITIES

    o   R CMD check now tests for mis-use on an installed or binary
    package, and sets 'T' and 'F' to 'NULL' when running the
    examples.

    o   New function SweaveSyntConv() converts between Sweave file
    syntaxes. RweaveLatex() now gets its prompt from options() and
    uses the text width as linebreak cutoff for deparsing input
    statements.


BUG FIXES

    o   axis() was not respecting par("mgp")[3] by default.
    (PR#916)

    o   tcltk back-compatibility fix for tcl8.0

    o   hist.POSIXct(*, breaks) now works for `breaks = #{breaks}' and
    when `x' has NAs; the latter applies to cut.POSIXct() as well.

    o   The internal download.file() methods were setting the proxy
    port to 0 unless it was specified.

    o   poly() did not work for >=3 column matrices.

    o   cut.default() was not handling infinite values with infinite
    breaks.  (PR#1694)

    o   ks.test() could fail due to integer overflow if n.x and n.y
    were both large.

    o   all.equal(3., 3:3) wasn't TRUE when the methods package was
    present.

    o   read.table() could remap non-syntactic names to duplicates,
    and did not check for duplicated names.

    o   ls.str(envir = environment(F)) now works (when F is a function).

    o   pie() now has a `border' (and 'lty') argument which allows empty
    slice borders and makes pie() useable for hundreds of slices.

    o   all.equal.character() now works correctly also when NAs don't
    match (PR#1767).

    o   all.equal.numeric() now gives character {instead of list} when
    lengths don't match.

    o   read.dcf() had a memory leak.

    o   Setting "tcl = a" in a highlevel graphic function worked almost
    as if par(tcl = a) was called (i.e. was persistent).  Further,
    after par(tcl= <.>), highlevel graphic setting of tcl didn't work
    anymore.  "tck" has now S's default of -0.01 (when tcl=NA).

    o   data() was not checking if ./data was a directory before
    warning.

    o   Assignment beyond the end of character strings was filling
    with "", whereas changing the length extended with NA_STRING.
    Now NA_STRING is used for both.

    o   boxplot()s with logarithmic scale in "width-direction" now have
    proper widths, i.e. typically constant.

    o   Using GNU readline in asynchronous event callbacks (e.g. TclTk
    or Gtk) works correctly, i.e. doesn't cause readline to abort
    the process.

    o   Using Q to exit a browser invoked at the top-level could lead to
    the browser being invoked on all future top-level loops (PR#1721)

    o   step(fit, direction="both") now uses both directions even if
    no scope is supplied.

    o   strwidth(), strheight() now give 0 on NA strings (as they
    are no longer plotted). (PR#1739)

    o   mtext(), persp(), plot.hclust() and title() no longer plot NA
    character strings.

    o   pdf() has been protected against the user who specified
    non-existent fonts. (PR#1748)

    o   dlnorm() returns zero for negative x argument. (PR#1781)

    o   Printing 0 in octmode gave "" not "0". (PR#1759)

    o   anova.glm() was ignoring a `dispersion' argument if given
    multiple objects. (PR#1807)

    o   model.matrix() with a `rhs ~ .' formula included columns with
    duplicated names in the data frame, which caused subsequent
    confusion (e.g. in prediction from the object).

    model.matrix() was sometimes incorrectly determining the first
    factor in a formula without an intercept.

    o   termplot() now needs data= argument more rarely. (PR#828)

    o   abline() on log scale doesn't draw -ve points. (PR#1243)

    o   The Java search engine now works within Mozilla 1.0 and
    Netscape 6.2.x / 7.0pr1.

    o   unsplit() failed if f was a list of factors. (PR#1843)

    o   The methods package generic version of primitives is now "sealed"
    and cannot be redefined (it was always a bad idea to do so).

    o   quantile() gave -Inf not NaN in some examples. (Related to PR#1852)

    o   read.table() read too far in checking the file structure if
    0 < nrows < 5 and more rows existed on the file. (PR1809)

    o   loess() was not checking for too small a span (so no points
    were included in the smoothing window).

    o   match() was assuming that there was only one possible bit
    pattern for a numeric NA, so some matches failed on Solaris
    under some compiler options.

    o   identical() no longer thinks NaN and as.double(NA) are identical.

    o   pipe() can open in a binary mode even on Linux (where popen
    cannot).

    o   zero-column matrices were not being printed at all, not even
    the row names.

    o   polygon()'s `border' argument was incorrectly documented: in
    particular `border = 0' plots in background colour.

    o   delete.response() was losing the attribute set for safe
    prediction. (PR#1840)

    o   poly() was checking the degree against the number of points
    even when predicting.

    o   Comparison of a data frame and list failed due to typo. (PR#1889)

    o   dput() and dump() attempt to check if writing succeeded. (PR#1884)

    o   rep(1:2, 0) and hence array(1:2, 0) fail no longer.

    o   apply(matrix("", 3, 0), 2, length) now works.

    o   order(na.last = NA, decreasing = TRUE) now sorts in decreasing
    order.  (PR#1906)

    o   order(na.last=NA) failed if all args were of length 1 or all
    were NA. (PR#1913, 1981)

    o   source() would crash if the `keep.source = TRUE' and the input
    contained a function with a line longer than 1024 chars. (PR#1900)

    The limits on nesting of functions and on total function size
    when the source is kept are now enforced.  As (incorrectly)
    the source was always kept, this meant functions longer than
    128Kb could crash R, even in packages.

    o   lqs() and cov.rob() check that the quantile argument does not
    exceed n-1.

    o   grid() now also works with log coordinates active.

    o   plot.table(tab, xlab="X", ylab="Y") now works as expected.

    o   plot.formula(ask=FALSE) now works. (PR#1923)

    o   as.list(<logical>) now works properly. (PR#1926).

    o   The residual SSq printed from an aov() fit with weights is
    now the weighted SSq.  (PR#1930)

    o   aov() could still fail on multistrata models with very long
    Error formulae.

    o   try() could fail in BATCH use due to use of fflush(stdin).
    Only seen on Solaris. (PR#1934)

   o    title(main=list("main")) printed garbage, as did similar calls
    using quote() instead of expression() for plotmath. (PR#1939)

   o    deparse() dropped parentheses in some case where they were
    needed. (PR#1119, 1737, 1928, at least)

   o    pdf(onefile = FALSE) never incremented the file number beyond
    two.

   o    On Unix, protect against broken pipes where popen calls succeed
    even though the command does not exist (and the glibc manual
    says it should return NULL).  (PR#1959)

   o    data.frame() was allowing explicit row.names with missing
    values, but row names are required to be unique.

   o    as.character(expression) was silently truncating to 60
    characters: the help page claimed 500 which is now true.

   o    as.hclust(x) now also works (as identity) for "hclust" objects.

   o    NextMethod didn't work right in Ops.

   o    dotchart() now obeys `xlim ='.

   o    t(x) behaves when x is a (multivariate) time series. (PR#1998)

   o    parse(text=x) no longer parses from stdin if length(x) is zero

   o    binom.test() miscalculated p-value in some extreme cases

   o    get("print.ts")(1) would segfault. It's now just an error.

   o    cbind(NULL) {and similar} gave an error instead of NULL.

   o    the complex version of solve(a,b) now also works if b is a
    vector


        CHANGES IN R VERSION 1.5.1


NEW FEATURES

    o   Enhancements to mathematical annotation of plots:

    - expressions involving dot(<something>) now produce a dot accent
      above the <something> (initial patch from Ben Bolker).

    - within an expression, the symbol partialdiff is now converted
      to a partial differential symbol (greek delta).

    o   smooth.spline() has a new argument `nknots' allowing to set the
    default number of knots (when `all.knots = FALSE' as per default).


BUG FIXES

    o   Rdconv now skips CRs in input files, even on Unix.

    o   readBin() had a (very slow) memory leak if changing size or reading
    character strings.  writeChar() had a memory leak.

    o   polygon() and rect() with lty = "blank" were omitting the fill
    colour (if present).  symbols(circles=, lty="blank") was
    drawing the border of the disks.

    o   Subsetting non-existent rows in a data frame would produce
    missing row names, which are not allowed.

    o   On Unix, R_PAPERSIZE was set to '' instead of 'a4' as default on
    systems without paperconf.

    o   Under GNOME, capabilities() now reports correctly that X11, png
    and jpeg are available (if they are under X11).

    o   The names of some results of unlist() had the wrong internal
    length, which confused paste().  Both have been corrected. (PR#1524)

    o   RweaveLatex.Rd had unbalanced braces in section "Supported
    Options".

    o   merge() with multiple and differently-named match columns
    failed. (PR#1510)

    o   NAs in right-justified unquoted character matrices were being
    mis-aligned.  This mainly affected printing data frames.

    o   predict.*bSpline() bugs extrapolating for deriv >= 1 (PR#1473),
    and predict.[bn]s bug e.g. for bs(x), reported by Ch.Sangiorgio.

    o   qr.X was failing if n < p.  Fixed, but only in the case when
    pivoting does not occur. (PR#1519)

    o   xx[, 1:3] was returning a list if xx had only one row, even
    though xx[1, 1:3] was a data frame. (PR#1530)

    o   nls() was reporting incorrectly the number of iterations if the
    maximum was reached.

    o   rbind.data.frame() was coercing logical columns to factors.
    (PR#1536)

    o   Rprof(NULL) or Rprof("") called when not profiling caused a
    segfault: now silently ignored.  (PR#1471)

    o   On systems (e.g. Windows) using R's own code for expm1, the
    values were wrong for large negative x, and this affected
    pweibull(), for example.

    o   prettyNum(*, bigmark=*) added extra marks in some cases. (PR#1548)

    o   cut.dendrogram() failed in some cases. (PR#1552)

    o   The links in refman.pdf were broken by a single invalid
    \alias{}. Now Rdconv checks more comprehensively. (PR#1550)

    o   predict(smooth.spline(x,*), deriv =1) did not predict at all `x'
    values, and
    smooth.spline(x,*, all.knots = TRUE) used much too much memory
    when length(x) was largish.
    smooth.spline(*, .... trace = TRUE ..) is a bit more self-explaining.

    o   pexp(x, log=TRUE) and pweibull(*) were losing precision for large x.

    o   Workaround for NetBSD bug in compiling dounzip.c. (PR#1565/6)

    o   Conversions to numeric in type.convert were always treating "NA" as
    missing whatever the setting of na.strings. (PR#1568)

    o   R CMD check now deals correctly with package names containing a
    '.' in the Depends field of a DESCRIPTION file. (PR#1591)

    o   pbinom() and pbinom() were misssing the "fuzz" that other discrete
    distributions have in order to guard against truncation.

    o   The "=" assignment operator behaved as "<<-" in some constructions
    with composite left-hand sides.

    o   Added print.AsIs() method to ensure "AsIs" objects are printed
    by the method for their real class. (PR#1587)

    o   mosaicplot(*, color = v) now recycles v[] if needed.

    o   Calling dev.control("inhibit") with no graphics device open
    crashed R. (PR#1605)

    o   read.dcf() was not checking that memory allocation succeeded.

    o   The C function fprec might overflow for > 16 digits, showing
    up in example(str) on some platforms.

    o   Changes to the methods package:

    - Nonstandard generic function definitions given to
      setGeneric now work; these were previously ignored.

    - Classes that extend matrix, etc. as the data part previously lost
      their attributes.

    - There were bugs in distinguishing methods for "missing" from
      methods for "ANY".

    - Some coerce methods were not selected correctly when the two classes
      were related by an "is" relation.

    - extends() now works correctly if given a class definition rather
      than a name.

    - class()<- was leaving in an explicit class attribute for basic data
      types.

    - Method signatures including arguments following "..." in the
      argument list of the generic didn't work.

    o   par(mgp = v) now allows negative `v' (S-compatibly and sensibly).

    o   identify.hclust() masked its `x' argument and so failed.

    o   System rint was never being used in fround.c.

    o   round/trunc.POSIXt were erroneously retaining information on
    DST. (PR#1543)

    o   tcltk package didn't work with tcl8.0.x.  Compatibility code
    inserted (PR#1640, thanks to René Bertin for helping out with
    this).

    o   selfStart() {pkg "nls"} now works again when parameter names are
    not specified (but implicit).

    o   legend() now treats lty="0" properly (as solid).

    o   ARMAacf was failing if 0 < p < q+1: it failed to say that ar needed
    to be padded with zeroes, and now the code does that.

    o   old.packages() was assuming a bundle was installed in only one
    library, and so update.packages() would only update a bundle
    in the first location in lib.loc.

    o   fisher.test() could miscalculate odds ratio and confidence
    interval for tables with large entries, due to numerical
    overflow.

    o   Using lty = "1" (or "10") does no longer produce invalid
    postscript or pdf.  Using "0" in a character lty is currently not
    implemented device-independently and needs more fixing.

    o   ccf(plot = FALSE) was returning a spurious extra 0 at lag 0.

    o   On console-based versions of R (Windows, GNOME, probably
    Mac OS) warning/error messages longer than 8192 chars could
    crash R. (PR#1651)

    o   Comparisons between objects of mode "call" using "==" now return
    TRUE if both sides deparse to the same string (as always intended).

    o   HoltWinters() now initializes correctly in the exponential
    smoothing case, and plot.HoltWinters() allows lty to be set.

    o   fisher.test() crashed due to corruption on some large
    problems.  It also crashed on tables with total one. (PR#1662)

    o   The substitution code for strptime (used e.g. on Windows)
    cached the month names and so did not recognise locale changes
    during an R session, as used by get.hist.quote{tseries}.
    Caching has been supressed.  (PR#1116)

    o   Some functions used the non-existing error() function instead
    of stop().

    o   vector("complex",n)  doesn't return random garbage anymore.


        CHANGES IN R VERSION 1.5.0


USER-VISIBLE CHANGES

    o   XDR support is now guaranteed to be available, so the default
    save format will always be XDR binary files, and it is safe to
    distribute data in that format.  (We are unaware of any
    platform that did not support XDR in recent versions of R.)

    gzfile() is guaranteed to be available, so the preferred
    method to distribute sizeable data objects is now via
    save(compress = TRUE).

    o   pie() replaces piechart() and defaults to using pastel colours.

    o   formatC has new arguments (see below) and formatC(*, d = <dig>)
    is no longer valid and must be written as formatC(*, digits = <dig>).

    o   Missingness of character strings is treated much more
    consistently, and the character string "NA" can be used as a
    non-missing value.

    o   summary.factor() now uses a stable sort, so the output will
    change where there are ties in the frequencies.


NEW FEATURES

    o   Changes in handling missing character strings:

    -  "NA" is no longer automatically coerced to a missing value
       for a character string.  Use as.character(NA) where a missing
       value is required, and test via is.na(x) not x == "NA".
       String "NA" is still converted to missing by scan() and
       read.table() unless `na.strings' is changed from the default.

    -  A missing character string is now printed as `NA' (no quotes)
       amongst quoted character strings, and `<NA>' if amongst
       unquoted character strings.

    -  axis() and text.default() omit missing values of their
       `labels' argument (rather than plotting "NA").

    -  Missing character strings are treated as missing much more
       consistently, e.g. in logical comparisons and in sorts.
       identical() now differentiates "NA" from the missing string.


    o   Changes in package methods:

    -  New function validSlotNames().

    -  Classes can explicitly have a "data part", formally
       represented as a .Data slot in the class definition, but
       implemented consistently with informal structures.  While the
       implementation is different, the user-level behavior largely
       follows the discussion in "Programming with Data".

       -   A "next method" facility has been provided, via the function
       callNextMethod().  This calls the method that would have been
       selected if the currently active method didn't exist.  See
       ?callNextMethod().  This is an extension to the API.

       -   Classes can have initialize methods, which will be called
       when the function new() is used to create an object from the
       class.  See ?initialize.  This is an extension to the API.

       -   The logic of setGeneric() has been clarified, simplifying
       nonstandard generic functions and default methods.


    o   Changes in package tcltk:

    -  Now works with the GNOME user interface.

    -  Several new functions allow access to C level Tcl objects.
       These are implemented using a new `tclObj' class, and this
       is now the class of the return value from .Tcl() and tkcmd().


    o   Changes in package ts:

    -  More emphasis on handling time series with missing values
       where possible, for example in acf() and in the
       ARIMA-fitting functions.

    -  New function arima() which will replace arima0() in due
       course.  Meanwhile, arima0() has been enhanced in several
       ways.  Missing values are accepted.  Parameter values can
       be initialized and can held fixed during fitting.  There is
       a new argument `method' giving the option to use
       conditional-sum-of-squares estimation.

    -  New function arima.sim().

    -  New datasets AirPassengers, Nile, UKgas and WWWusage,
       and a expanded version of UKDriverDeaths (as a multiple
       time series Seatbelts).

    -  New generic function tsdiag() and methods for arima and arima0,
       to produce diagnostic plots.  Supersedes arima0.diag().

    -  New functions ARMAacf() and ARMAtoMA() to compute
       theoretical quantities for an ARMA process.

    -  New function acf2AR() to compute the AR process with a given
       autocorrelation function.

    -  New function StructTS() to fit structural time series, and
       new generic function tsSmooth() for fixed-interval
       state-space smoothing of such models.

    -  New function monthplot() (contributed by Duncan Murdoch).

    -  New functions decompose() and HoltWinters() (contributed by
       David Meyer) for classical seasonal decomposition and
       exponentially-weighted forecasting.


    o   An extensible approach to safe prediction for models with e.g.
    poly(), bs() or ns() terms, using the new generic function
    makepredictcall().  Used by most model-fitting functions
    including lm() and glm().  See ?poly, ?cars and ?ns for
    examples.

    o   acosh(), asinh(), atanh() are guaranteed to be available.

    o   axis() now omits labels which are NA (but still draws the
    tick mark.

    o   Connections to bzip2-ed files via bzfile().

    o   chol() allows pivoting via new argument `pivot'.

    o   cmdscale() now takes rownames from a dist object `d' as well
    as from a matrix; it has new arguments `add' (as S) and `x.ret'.

    o   crossprod() handles the case of real matrices with y = x
    separately (by accepting y = NULL).  This gives a small
    performance gain (suggestion of Jonathan Rougier).

    o   deriv/deriv3() can now handle expressions involving pnorm and
    dnorm (with a single argument), as in S-PLUS.

    o   New function expm1() both in R and in C API, for accurate exp(x)-1;
    precision improvement in pexp() and pweibull() in some cases.
    (PR#1334-5)

    o   New function findInterval() {using new C entry point
    findInterval, see below}.

    o   formatDL() now also works if both items and descriptions are
    given in a suitable list or matrix.

    o   gzfile() is guaranteed to be available, and hence the
    `compress' option to save() and save.image().

    o   hist() now has a method for date-time objects.

    o   library() now checks the dependence on R version (if any) and
    warns if the package was built under a later version of R.

    o   library(help = PKG) now also returns the information about the
    package PKG.

    o   Added function logb(), same as log() but for S-PLUS
    compatibility (where log now has only one argument).

    o   New na.action function na.pass() passes through NAs unaltered.

    o   piechart() has been renamed to pie(), as piechart is a Trellis
    function for arrays of pie charts.  The default fill colours
    are now a set of pastel shades, rather than par("bg").

    o   plclust() in package mva, for more S-PLUS compatibility.

    o   poly() now works with more than one vector or a matrix as input,
    and has a predict method for objects created from a single
    vector.

    o   polyroot() now handles coefficient vectors with terminal
    zeroes (as in S).

    o   New prettyNum() function used in formatC() and format.default()
    which have new optional arguments `big.mark', `big.interval',
    `small.mark', `small.interval', and `decimal.mark'.

    o   print.coefmat() has a new argument 'eps.Pvalue' for determining
    when small P-values should be printed as  "< {...}".

    o   The recover() function has been moved to the base package.
    This is an interactive debugging function, usually a good
    choice for options(error=).  See ?recover.

    o   rep() has a new argument `each' for S-PLUS compatibility.
    The internal call is made available as rep.int(), again for
    help in porting code.

    o   New functions rowSums(), colSums(), rowMeans() and colMeans():
    versions of apply() optimized for these cases.

    o   rug() now has a "..." argument allowing its location to be specified.

    o   scan() can have NULL elements in `what', useful to save space
    when columns need to be discarded.

    o   New option by = "DSTday" for seq.POSIXt().

    o   Changes to sorting:

    -  sort(), sort.list() and order() have a new argument
       `decreasing' to allow the order to be reversed whilst
       still preserving ties.

    -  sort() has an option to use quicksort in some cases
       (currently numeric vectors and increasing order).

    -  The default Shell sort is Sedgewick's variant, around 20%
       faster, and pre-screening for NAs speeds cases without any
       NAs several-fold.

    -  sort.list() (and order with just one vector) is several
       times faster for numeric, integer and logical vectors, and
       faster for character vectors.

    o   New assignment forms of split(); new function unsplit().

    o   New sprintf() function for general C like formatting, from
    Jonathan Rougier.

    o   Argument `split' of summary.{aov,aovlist} is now implemented.

    o   summary.princomp() now has a separate print() method, and
    `digits' is now an argument to the print method and not to
    summary.princomp itself.

    o   An extended version of the trace() function is available,
    compatible with the function in S-PLUS.  Calls to R functions
    can be inserted on entry, on exit, and before any
    subexpressions. Calls to browser() and recover() are useful.
    See ?trace.

    o   New function TukeyHSD() for multiple comparisons in the results
    of aov().  (Formerly function Tukey in package Devore5 by
    Douglas Bates.)

    o   New read-only connections to files in zip files via unz().

    o   warning() has new argument `call.', like stop()'s.

    o   zip.file.extract() is no longer provisional and has an
    "internal" method available on all platforms.

    o   Methods for [, [<- and as.data.frame() for class "POSIXlt".

    o   Much improved printing of matrices and arrays of type "list".

    o   The "Knuth-TAOCP" option for random-number generation has been
    given an option of using the 2002 revision.  See ?RNG for the
    details: the R usage already protected against the
    reported `weakness'.

    o   min/max of integer(0) (or NULL) is now Inf/-Inf, not an
    extreme integer.


DEPRECATED & DEFUNCT

    o   .Alias, reshapeLong(), reshapeWide() are defunct.

    o   arima0.diag() (package ts) is deprecated: use tsdiag() instead.

    o   piechart() is deprecated; renamed to pie().


DOCUMENTATION CHANGES

    o   `Writing R Extensions' now has an example of calling R's
    random numbers from FORTRAN via C.

    o   R itself and all R manuals now have ISBN numbers, please use
    them when citing R or one of the manuals.


INSTALLATION CHANGES

    o   The configure script used when building R from source under Unix
    is now generated using Autoconf 2.50 or later, which has the
    following 'visible' consequences:

    - By default, configure no longer uses a cache file.  Use the
      command line option '--config-cache' (or '-C') to enable
      caching.

    - Key configuration variables such as 'CC' are now *precious*,
      implying that the variables

      * no longer need to be exported to the environment and can and
        should be set as command line arguments;

      * are kept in the cache even if not specified on the command
        line, and checked for consistency between two configure runs
        (provided that caching is used, see above);

      * are kept during automatic reconfiguration as if having been
        passed as command line arguments, even if no cache is used.

      See the variable output section of 'configure --help' for a
      list of all these variables.

    o   Configure variable 'FC' is deprecated, and options '--with-g77',
    '--with-f77' and '--with-f2c' are defunct.  Use configure
    variable 'F77' to specify the FORTRAN 77 compiler, and 'F2C' to
    specify the FORTRAN-to-C compiler and/or that it should be used
    even if a FORTRAN 77 compiler is available.

    o   Non-standard directories containing libraries are specified
    using configure variable 'LDFLAGS' (not 'LIBS').


UTILITIES

    o   Sweave(), Stangle() and friends in package tools.  Sweave allows
    mixing LaTeX documentation and R code in a single source file:
    the R code can be replaced by its output (text, figures) to
    allow automatic report generation.  Sweave files found in
    package subdir 'inst/doc' are automatically tested by R CMD
    check and converted to PDF by R CMD build, see the section on
    package vignettes in the 'Writing R Extensions' manual.

    o   Rdconv can convert to the S4 '.sgml' format.

    o   R::Utils.pm masks some platform dependencies in perl code by
    providing global variables like R_OSTYPE or wrapper functions
    like R_runR().

    o   If a directory 'inst/doc' is present in the sources of a
    package, the HTML index of the installed package has a link to
    the respective subdirectory.

    o   R CMD check is more stringent: it now also fails on malformed
    'Depends' and 'Maintainer' fields in 'DESCRIPTION' files, and on
    unbalanced braces in Rd files.  It now also provides pointers to
    documentation for problems it reports.

    o   R CMD check, build and INSTALL produce outline-type output.

    o   QC functions in package 'tools' now return the results of their
    computations as objects with suitable print() methods.  By
    default, output is only produced if a problem was found.

    o   New utility R CMD config to get the values of basic R configure
    variables, or the header and library flags necessary for linking
    against R.

    o   Rdindex and maketitle.pl require perl 5.005, as Text::Wrap::fill
    was only introduced at 5.004_05.


C-LEVEL FACILITIES

    o   All the double-precision BLAS routines are now available,
    and package writers are encouraged not to include their own
    (so enhanced ones will be used if requested at configuration).

    o   findInterval(xt[],n,x,...) gives the index (or interval number)
    of x in the sorted sequence xt[].  There's an
    F77_SUB(interv)(.) to be called from Fortran; this used to be
    part of predict.smooth.spline's underlying Fortran code.

    o   Substitutes for (v)snprintf will be used if the OS does not
    supply one, so tests for HAVE_(V)SNPRINTF are no longer needed.

    o   The DUP and NAOK arguments in a .C() call are not passed on
    to the native routine being invoked.  Any code that relied on
    the old behaviour will need to be modified.

    o   log1p is only provided in Rmath.h if it is not provided by the
    platform, in which case its name is not remapped, but a
    back-compatibility entry point Rf_log1p is provided.
    Applications using libRmath may need to be re-compiled.

    o   The methods used by optim() and integrate() have entry points in
    R_ext/Applic.h and have a more general interface documented in
    `Writing R Extensions'.

    o   The bessel_? entry points are now suitable to be called
    repeatedly from code loaded by .C().  (They did not free
    memory until .C() returned in earlier versions of R.)

    o   Server sockets on non-Windows platforms now set the SO_REUSEADDR
    socket option.  This allows a server to create simultanous
    connections to several clients.

    o   New quicksort sorting (for numeric no-NA data), accessible from
    C as R_qsort() etc and from Fortran as qsort4() and qsort3().

    o   Rinternals.h no longer includes fcntl.h, as this is not an
    ISO C header and cannot be guaranteed to exist.

    o   Fortran subroutines are more correctly declared as `extern void'
    in R_exts/Applic.h and R_exts/Linpack.h.


BUG FIXES

    o   The calculation of which axes to label on a persp() plot was
    incorrect in some cases.

    o   Insufficient information was being recorded in the display list
    for the identify() function.  In particular, the "plot="
    argument was ignored when replaying the display list. (PR#1157)

    o   The vertical alignment of mathematical annotations was wrong.
    When a vertical adjustment was not given, it was bottom-adjusting
    i.e,. it was treating adj=0 as adj=c(0, 0).  It now treats
    adj=0 as adj=c(0, 0.5) as for "normal" text. (PR#1302)

    o   the man page (doc/R.1) wasn't updated with the proper VERSION.

    o   smooth.spline() had a "df = 5" default which was never used and
    hence extraneous and misleading.

    o   read.fwf() was interpreting comment chars in its call to scan:
    replaced by a call to readlines(). (PR#1297/8)

    o   The default has been changed to scan(comment.char="") for
    consistency with earlier code (as in the previous item).

    o   bxp(*, notch.frac = f) now draws the median line correctly.

    o   Current versions of gs were rotating the output of
    bitmap(type = "pdfwrite") and when converting the output
    of postscript() to PDF; this has been circumvented by
    suppressing the %%Orientation comment for non-standard
    paper sizes.

    o   plot.ts(x, log = "y") works again when x has 0s, also for matrix x.

    o   add1(), drop1(), step() work again on glm objects with
    formulae with rhs's containing `.'.  (Broken by a `bug fix'
    (in reality an API change) in 1.2.1.)

    o   optim(method="BFGS") was not reporting reaching `maxit'
    iterations in the convergence component of the return value.

    o   aov() and model.tables() were failing on multistrata models with
    excessively long Error formula. (PR#1315)

    o   Transparent backgrounds on png() devices on Unix-alikes had
    been broken during the driver changes just prior to 1.4.0.
    (They worked correctly on Windows.)

    o   demo(is.things) didn't work properly when the methods package was
    attached.

    o   match(), unique() and duplicated() were not declaring all NaNs to be
    equal, yet not always distinguishing NA and NaN.  This was
    very rare except for data imported as binary numbers.

    o   The error handler recover() protects itself against errors in
    dump.frames and uses a new utility, limitedLabels, to generate
    names for the dump that don't inadvertently blow the limit on
    symbol length.  (TODO:  either fix dump.frames accordingly or
    remove the limit--say by truncating very long symbols?)

    o   se.contrasts() works more reliably with multistratum models,
    and its help page has an example.

    o   summary.lm() was not returning r.squared nor adj.r.squared for
    intercept-only models, but summary.lm.null() was returning
    r.squared but not adj.r.squared.  Now both are always
    returned.  Neither returned f.statistic, and that is now
    documented.

    o   Subsetting of matrices of mode list (or other non-atomic
    modes) was not implemented and gave incorrect results without
    warning. (PR#1329)
    Under some circumstances subsetting of a character matrix
    inserted NA in the wrong place.

    o   abs() was not being treated as member of the Math group
    generic function, so e.g. its method for data frames was not
    being used.

    o   set.seed(seed, "default") was not using the `seed' value (only
    for kind = "default").

    o   logLik.lm() now uses  "df = p + 1" again (`+ sigma'!).

    o   logLik.glm() was incorrect for families with estimated dispersion.

    o   Added strptime() workaround for those platforms (such as
    Solaris) that returned missing components as 0.  Missing
    days are now detected, but missing years will still
    be interpreted as 1900 on such platforms.

    o   Inheritance in formal classes (the methods package) works
    breadth-first as intuition would expect.

    o   The new() function in package `methods' works better (maybe even
    correctly?) for the various combinations of super-classes and
    prototypes that can be supplied as unnamed arguments.

    o   Internal code allowed one more connection to be allocated than
    the table size, leading to segfaults. (PR#1333)

    o   If a user asks to open a connection when it is created and
    it cannot be opened, the connection is destroyed before returning
    from the creation call.  (related to PR#1333)

    o   Sys.putenv() was not using permanent storage. (PR#1371)

    o   La.svd() was not coercing integer matrices. (PR#1363)

    o   deriv(3) now reports correctly the function it cannot find
    the derivatives table.

    o   The GNOME user interface was over-enthusiastic about setting
    locale information.  Now only LC_CTYPE, LC_COLLATE and LC_TIME
    are determined by the user's environment variables (PR#1321)

    o   In X11, locator() would sound the bell even if "xset b off"
    had been set.

    o   merge() could be confused by inconsistent use of as.character()
    giving leading spaces.

    o   [pqr]binom() no longer silently round the `size' argument,
    but return NaN (as dbinom() does). (PR#1377)

    o   Fixed socket writing code to block until all data is written.
    Fixed socket reading code to properly handle long reads and reads
    with part of the data in the connection buffer.

    o   Allow sockets to be opened in binary mode with both open="ab" and
    open="a+b".

    o   levels<-.factor() was using incorrectly list values longer than
    the number of levels (PR#1394), and incorrectly documented
    that a character value could not be longer than the existing
    levels.

    o   The pdf() device was running out of objects before the
    documented 500 page limit.  Now there is no limit.

    o   legend() did not deal correctly with `angle' arguments. (PR#1404)

    o   sum() tried to give an integer result for integer arguments,
    but (PR#1408)
    - this was not documented
    - it sometimes warned on overflow, sometimes not
    - it was order-dependent for a mixture of integer and numeric args.

    o   mean() gave (numeric) NA if integer overflow occurred in sum(),
    but now always works internally with numeric (or complex) numbers.

    o   sort.list() and order() were treating NA_STRING as "NA".

    o   sort.list(na.last = NA) was not implemented.

    o   seq.default() was returning only one element for a relative
    range of less than about 1e-8, which was excessively conservative.
    (PR#1416)

    o   tsp(x) <- NULL  now also works after library(methods).

    o   persp(shade=) was not working correctly with the default
    col=NULL if this was transparent. (PR#1419)

    o   min/max(complex(0)) was returning a random value.

    o   range() gave c(1, 1).

    o   range(numeric(0)) is now c(Inf, -Inf), as it was documented to be.

    o   print.ts() was occasionally making rounding errors in the labels
    for multiple calendar time series.

    o   Rdconv was not handling nested \describe{} constructs when
    converting to HTML (PR#1257) and not fixing up mal-formed
    \item fields in \describe{} when converting to text (PR#1330).

    o   filled.contour() was not checking consistency of x, y, z.
    (PR#1432)

    o   persp.default() no longer crashes with non-character labels.
    (PR#1431)

    o   fft() gave incorrect answers for input sizes 392, 588, 968, 980 ....
    (PR#1429)

    o   det(method = "qr") gave incorrect results for numerically
    singular matrices. (PR#1244)

    o   barplot() now allows the user to control `xpd'. (PR#1088, 1398)

    o   library() (with no arguments) no longer fails on empty TITLE files.

    o   glm() was failing if both offset() and start were specified. (PR#1421)

    o   glm() might have gotten confused if both step-shortening and
    pivoting had occurred (PR#1331).  Step-halving to avoid the
    boundary of feasible values was not working.

    o   Internal representation of logical values was not being
    treated consistently. (Related to PR#1439)

    o   The c() function sometimes inserted garbage in the name vector
    for some types of objects, e.g.  names(c(ls, a=1)).

    o   Fixed bug in `$' that could cause mutations on assignment (PR#1450).

    o   Some X servers displayed random bytes in the window title of
    graphics windows (PR#1451)

    o   The X11 data editor would segfault if closed with window manager
    controls (PR#1453)

    o   Interrupt of Sys.sleep() on UNIX no longer causes subsequent
    Sys.sleep() calls to segfault due to infinite recusion.

    o   Eliminated a race condition that could cause segfaults when a SIGINT
    was received while handling an earlier SIGINT.

    o   rect(lty = "blank") was incorrectly drawing with a dashed line.

    o   type.convert() was not reporting incorrectly formatted complex
    inputs. (PR#1477)

    o   readChar() was not resetting vmax, so causing memory build-up.
    (PR#1483)



    **************************************************
    *                        *
    *          1.4 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 1.4.1


BUG FIXES

    o   scan(multi.line = FALSE) now always gives an immediate error
    message if a line is incomplete. (As requested in PR#1210)

    o   read.table() is no longer very slow in processing comments:
    moved to C code and fewer lines checked.

    o   type.convert() could give stack imbalance warnings if used
    with as.is = TRUE.

    o   predict.mlm ignored newdata (PR#1226) and also offsets.

    o   demo(tkttest) was inadvertently changed in 1.4.0 so that it would
    evaluate the requested test, but not display the result.

    o   stars(scale = TRUE) (the default) now works as documented (and
    as S does).  Previously it only scaled the maximum to 1. (PR#1230)

    o   d0 <- data.frame(a = 0); data.matrix(d0[0, 0]) and
    data.matrix(d0[, 0]) now work.

    o   plot(multiple time series, plot.type = "single") was computing
    `ylim' from the first series only.

    o   plot.acf() has a new `xpd = par("xpd")' argument which by default
    *does* clipping (of the horizontal lines) as desired (xpd = NA was
    used before, erronously in most cases).

    o   predict(smooth.spline(.), deriv = 1) now works.

    o   identify() failed when x is a structure/matrix. (PR#1238)

    o   getMethod() returns NULL when optional=TRUE as promised in the
    documentation.

    o   setMethod() allows "..." to be one of the arguments omitted in
    the method definition (but so far no check for ... being missing)

    o   Allow round() to work again on very large numbers (introduced
    in fixing PR#1138). (PR#1254)

    o   Rinternals.h is now accepted by a C++ compiler.

    o   type.convert() was failing to detect integer overflow.

    o   piechart() was defaulting to foreground colour (black) fills
    rather than background (as used in 1.3.1 and earlier).  Now
    background is used, but be aware that as from 1.4.0 this may
    be transparent.

    o   La.eigen(*, only.values=TRUE) does not segfault anymore in one
    branch (PR#1262).

    o   cut() now produces correct default labels even when
    include.lowest = TRUE (PR#1263).

    o   reformulate() works properly with a response.

    o   cmdscale(*, k = 1) now works properly.

    o   Options by = "month" and "year" to seq.POSIXt() will always
    take account of changes to/from daylight savings time: this
    was not working on some platforms.

    o   glm.fit.null() now accepts all the arguments of glm.fit() (it
    could be called from glm.fit with arguments it did not
    accept), and is now documented.

    o   cov.wt(cbind(1), cor = TRUE) now works.

    o   predict(glm.object, se.fit = TRUE) was failing if the fit
    involved an offset.

    o   detach() on package:base would crash R. (PR#1271)

    o   print or summary on a manova() object with no terms, no names on
    the response and intercept = FALSE (which is not sensible)
    would give an error.

    o   seek() on file connections was ignoring the `origin' argument.

    o   Fixed new environment handling in library() to avoid forcing
    promises created by delay()

    o   arima0() could leak memory: now released via on.exit().

    o   qr.coef(qr,*) now keeps the names of qr$qr.

    o   read.00Index() no longer fails on data indexes not generated by
    Rdindex (PR#1274).


        CHANGES IN R VERSION 1.4.0


USER-VISIBLE CHANGES

    [This is a new section to highlight changes in behaviour, which
    may be given in more detail in the following sections.
    Many bug fixes are also user-visible changes.]

    o   The default save format has been changed, so saved workspaces
    and objects cannot (by default) be read in earlier versions of R.

    o   The number of bins selected by default in a histogram uses
    the correct version of Sturges' formula and will usually be
    one larger.

    o   data.frame() no longer converts logical arguments to factors
    (following S4 rather than S3).

    o   read.table() has new arguments `nrows' and `colClasses'.  If the
    latter is NA (the default), conversion is attempted to
    logical, integer, numeric or complex, not just to numeric.

    o   model.matrix() treats logical variables as a factors with
    levels c(FALSE, TRUE) (rather than 0-1 valued numerical
    variables).  This makes R compatible with all S versions.

    o   Transparency is now supported on most graphics devices.
    This means that using par("bg"), for example in legend(), will
    by default give a transparent rather than opaque background.

    o   [dpqr]gamma now has third argument `rate' for S-compatibility
    (and for compatibility with exponentials).  Calls which use
    positional matching may need to be altered.

    o   The meaning of spar = 0 in smooth.spline() has changed.

    o   substring() and substring()<- do nothing silently on a
    character vector of length 0, rather than generating an
    error.  This is consistent with other functions and with S.

    o   For compatibility with S4, any arithmetic operation using a
    zero-length vector has a zero-length result.  (This was
    already true for logical operations, which were compatible
    with S4 rather than S3.)

    o   undoc() and codoc() have been moved to the new package `tools'.

    o   The name of the site profile now defaults to
    `R_HOME/etc/Rprofile.site'.

    o   The startup process for setting environment variables now first
    searches for a site environment file (given by the environment
    variable `R_ENVIRON' if set or `R_HOME/etc/Renviron.site' if
    not), *and* then for a user `.Renviron' file in the current or
    the user's home directory.

    o   Former stars(*, colors = v) must now be stars(*, col.segments = v).

    o   The default methods for La.svd and La.eigen have changed and
    so there may be sign changes in singular/eigen vectors,
    including in cancor, cmdscale, factanal, princomp and varimax.


NEW FEATURES

    o   Transparency is now supported on most graphics devices.
    Internally colors include an alpha channel for opacity, but at
    present there is only visible support for transparent/opaque.
    The new color "transparent" (or NA or "NA") is transparent,
    and is the default background color for most devices.  Those
    devices (postscript, XFig, PDF, Windows metafile and printer)
    that previously treated bg = "white" as transparent now have
    "transparent" as the default and will actually print "white".
    (NB: you may have bg = "white" saved in .Postscript.options in
    your workspace.)

    o   A package `methods' has been added, containing formal classes
    and methods ("S4" methods), implementing the description in
    the book "Programming with Data".  See "?Methods" and the
    references there for more information.

    - In support of this, the "@" operator has been added to the
      grammar.

    - Method dispatch for formal methods (the standardGeneric
      function), is now a primitive. Aside from efficiency issues,
      this allows S3-style generics to also have formal methods
      (not really recommended in the long run, but it should at
      least work).  The C-level dispatch is now implemented for
      primitives that use either DispatchGroup or DispatchOrEval
      internally.

    - A version of the function "plot" in the methods package has
      arguments "x" and "y", to allow methods for either or both.
      See "?setMethod" for examples of such methods.

    - The methods package now uses C-level code (from within
      DispatchOrEval) to dispatch any methods defined for
      primitive functions.  As with S3-style methods, methods can
      only be defined if the first argument satisfies is.object(x)
      (not strictly required for formal methods, but imposed for
      now for simplicity and efficiency).

    o   Changes to the tcltk package:

    - New interface for accessing Tcl variables, effectively
      making the R representations lexically scoped. The old form
      is being deprecated.

    - Callbacks can now be expressions, with slightly unorthodox
      semantics. In particular this allows bindings to contain
      "break" expressions (this is necessary to bind code to
      e.g. Alt-x without having the key combination also insert an
      "x" in a text widget.)

    - A bunch of file handling and dialog functions (previously
      only available via tkcmd) have been added

    o   The "?" operator is now an actual function.  It can be used
    (as always) as a unary operator (?plot) and the grammar now
    allows it as a binary operator, planned to allow
    differentiating documentation on the same name but different
    type (class?matrix, for example).  So far, no such
    documentation exists.

    o   New methods AIC.default() and logLik.glm(), also fixing
    AIC(<glm obj>).

    o   axis.POSIXct() allows the label date/times to be specified
    via the new `at' argument.

    o   arrows() now allows length = 0 (and draws no arrowheads).

    o   Modifications to the access functions for more consistency with S:
    arguments `name', `pos' and `where' are more flexible in
    assign(), exists(), get(), ls(), objects(), remove() and rm().

    o   Three new primitive functions have been added to base:
    dataClass(), objWithClass(), and as.environment().  The first two
    are support routines for class() and class<-() in package
    methods.  The third replaces pos.to.env() in the functions get(),
    exists(), and friends.

    o   barplot() now respects an inline `cex.axis' argument and has a
    separate `cex.names' argument so names and the numeric axis
    labels can be scaled separately.  Also, graphics parameters
    intended for axis() such as `las' can now be used.

    o   Shading by lines added to functions barplot(), hist(),
    legend(), piechart(), polygon() and rect().

    o   bxp() has a show.names argument allowing labels on a single
    boxplot; it and hence boxplot() now makes use of pch, cex, and bg
    for outlier points().

    bxp() and boxplot() also have an argument `outline' to suppress
    outlier drawing {S-PLUS compatibly}.

    o   New capabilities() options "GNOME" and "IEEE754".

    o   New function casefold(), a wrapper for tolower/toupper
    provided for compatibility with S-PLUS.

    o   contour() is now generic.

    o   cor.test() in package ctest now also gives an asymptotic
    confidence interval for the Pearson product moment correlation
    coefficient.

    o   data(), demo() and library() now also return the information
    about available data sets, demos or packages.  Similarly,
    help.search() returns its results.

    o   density() allows `bw' or `width' to specify a rule to choose the
    bandwidth, and rules "nrd0" (the previous default), "nrd",
    "ucv", "bcv", "SJ-ste" and "SJ-dpi" are supplied (based on
    functions in package MASS).

    o   df.residual() now has a default method, used for classes
    "lm" and "glm".

    o   New argument `cacheOK' to download.file() to request cache
    flushing.

    All methods for download.file() do tilde-expansion on the path
    name.

    The internal download.file() etc now allow URLs of the form
    ftp://user@foo.bar/ and ftp://user:pass@foo.bar/

    o   duplicated() and unique() are now generic functions with
    methods for data frames (as well as atomic vectors).

    o   factanal() and princomp() use napredict() on their scores, so
    na.action = na.exclude is supported.

    o   Function getNativeSymbolInfo() returns details about a native
    routine, potentially including its address, the library in
    which it is located, the interface by which it can be called
    and the number of parameters.

    o   Functions such as help() which perform library or package index
    searches now use NULL as default for their `lib.loc' argument so
    that missingness can be propagated more easily.  The default
    corresponds to all currently known libraries as before.

    o   Added function file.rename().

    o   hist.default() allows `breaks' to specify a rule to choose the
    number of classes, and rules "Sturges" (the previous default),
    "Scott" and "FD" (Freedman-Diaconis) are supplied (based on
    package MASS).

    o   Function identical(), a fast and reliable way to test for exact
    equality of two objects.

    o   New generic function is.na<-(), from S4.  This is by default
    equivalent to x[value] <- NA but may differ, e.g. for factors
    where "NA" is a level.

    o   is.xxx reached through do_is are now generic.

    o   La.eigen() and La.svd() have new default methods to use later
    (and often much faster) LAPACK routines.  The difference
    is most noticeable on systems with optimized BLAS libraries.

    o   length() is now generic.

    o   New function .libPaths() for getting or setting the paths to the
    library trees R knows about.  This is still stored in .lib.loc,
    which however should no longer be accessed directly.

    o   Using lm/glm/... with `data' a matrix rather than a data frame
    now gives a specific error message.

    o   loess(), lqs(), nls() and ppr() use the standard NA-handling
    and so support na.action = na.exclude.

    o   mahalanobis() now has a `tol' argument to be passed to solve().

    o   mean() has `data frame' method applying mean column-by-column.
    When applied to non-numeric data mean() now returns NA rather
    than a confusing error message (for compatibility with S4).
    Logicals are still coerced to numeric.

    o   The formula interface to mosaicplot() now allows a contingency
    table as data argument.

    o   new.env() is now internal and allows you to set hashing. Also,
    parent.env() and parent.env<-() are included to provide direct
    access to setting and retrieving environments.

    o   Function nsl() to look up IP addresses of hosts: intended as
    a way to test for internet connectivity.

    o   Ops(), cbind(), diff() and na.omit() methods for time series
    objects moved from package ts to package base.

    o   New option `download.file.method' can be used to set the
    default method for download.file() and functions which use it
    such as update.packages().

    o   order() and sort.list() now implement na.last = FALSE, NA.

    o   Started work on new package management system: packageStatus()
    and friends.

    o   page() has a new `method' argument allowing `method = print'.

    o   png(), jpeg() and bmp() devices now have a `bg' argument to
    set the background color: useful to set "transparent" on png().

    o   Changes to the postscript() device:

    - The symbol font can now be set on a postscript() device, and
      support has been added for using Computer Modern type-1
      fonts (including for symbols). (Contributed by Brian D'Urso.)

    - There is now support for URW font families: this will give
      access to more characters and more appropriate metrics on
      PostScript devices using URW fonts (such as ghostscript).

    - %%IncludeResource comments have been added to the output.
      (Contributed by Brian D'Urso.)

    o   predict.ppr() now predicts on `newdata' containing NAs.

    o   princomp() now has a formula interface.

    o   readChar() now returns what is available if fewer characters
    than requested are on the file.

    o   readline() allows up to 256 chars for the prompt.

    o   read.table(), scan() and count.fields() have a new argument
    `comment.char', default `#', that can be used to start
    comments on a line.

    o   New function reg.finalizer() to provide R interface to
    finalization.

    o   reshape() extends reshapeLong, reshapeWide, which are deprecated.

    o   rle() now returns a classed object, has a print method and
    an inverse.

    o   Changes to save() and friends:

    - save() now takes an envir argument for specifying where
      items to be saved are to be found.

    - A new default format for saved workspaces has been
      introduced.  This format provides support for some new
      internal data types, produces smaller save files when saving
      code, and provides a basis for a more flexible serialization
      mechanism.

    - Modified `save' internals to improve performance when saving
      large collections of code.

    - save() and save.image() now take a `version' argument to
      specify the workspace file-format version to use.  The
      version used from R 0.99.0 to 1.3.1 is version 1.  The new
      default format is version 2.  load() can read a version 2
      saved workspace if it is compressed.

    - save() and save.image() now take a `compress' argument to
      specify that the saved image should be written using the zlib
      compression facilities.

    - save.image() now takes an argument `ascii'.

    - save.image() now takes an argument `safe'.  If TRUE, the
      default, a temporary file is used for creating the saved
      workspace.  The temporary file is renamed if the save
      succeeds.  This preserves an existing workspace if the save
      fails, but at the cost of using extra disk space during the
      save.

    - save.image() default arguments can be specified in the
      `save.image.defaults' option.  These specifications are used
      when save.image() is called from q() or GUI analogs.


    o   scan() allows unlimited (by R) lengths of input lines, instead
    of a limit of 8190 chars.

    o   smooth.spline() has a new `control.spar' argument and returns
    `lambda' besides `spar'.  spar <= 0 is now valid and allows
    to go more closely towards interpolation (lambda -> 0) than before.
    This also fixes smooth.spline() behavior for "df ~= n - 2". Better
    error messages in several situations.

    Note that spar = 0 is no longer the default and no longer entails
    cross-validation.

    o   stars() has been enhanced; new `mar' argument uses smaller
    mar(gins) by default; further `nrow and `ncol' as S-PLUS,
    `frame.plot', `flip.labels', `lty' and explicit `main', `sub',
    `xlab' and `ylab'.  Note that `colors' has been replaced by
    `col.segments' and there's a new `col.stars'.

    stars() now returns the locations invisibly.

    o   step() is now closer to stepAIC() and so handles a wider range
    of objects (but stepAIC [in MASS] is still more powerful).

    o   symbols() now has automatic xlab and ylab and a main argument
    which eliminates an incorrect warning.  It better checks wrongly
    scaled arguments.

    o   Sys.setlocale() now issues a warning if it fails.

    o   An enhanced function type.convert() is now a documented
    function, rather than just internal to read.table().

    o   warning() allows multiple arguments, following S4's style.

    o   New function with() for evaluating expressions in environments
    constructed from data.

    o   Unix x11() devices can now have a canvas color set, which can
    help to distinguish plotting "white" from plotting
    "transparent".

    o   On Unix, X11(), png() and jpeg() now give informative warnings
    if they fail to open the device.


    o   The startup processing now interprets escapes in the values of
    environment variables set in R_HOME/etc/Renviron in a similar
    way to most shells.

    o   The operator "=" is now allowed as an assignment
    operator in the grammar, for consistency with other languages,
    including recent versions of S-PLUS.  Assignments with "=" are
    basically allowed only at top-level and in braced or
    parenthesized expressions, to make famous errors such as
    "if(x=0) 1 else 2" illegal in the grammar.
    (There is a plan to gradually eliminate the underscore as an
    assignment in future versions of R.)

    o   Finalizers can be registered to be run on system exit for both
    reachable and unreachable objects.

    o   integer addition, subtraction, and multiplication now return NA's
    on overflow and issue a warning.

    o   Printing factors with both level "NA" and missing values
    uses `<NA>' for the missing values to distinguish them.

    o   Added an experimental interface for locking environments and
    individual bindings.  Also added support for "active bindings"
    that link a variable to a function (useful for example for linking
    an R variable to an internal C global).

    o   GNOME interface now has separate colours for input and output
    text (like the windows GUI). These can be modified via the
    properties dialogue.

    o   Output from the GNOME console is block buffered for increased
    speed

    o   The GNOME console inherits standard emacs-style keyboard
    shortcuts from the GtkText widget for cursor motion, editing
    and selection.  These have been modified to allow for the prompt
    at the beginning of the command line.

    o   One can register R functions and C routines to be called at the
    end of the successful evaluation of each top-level expression,
    for example to perform auto-saves, update displays, etc. See
    addTaskCallback() and taskCallbackManager(). See
    http://developer.r-project.org/TaskHandlers.pdf.


DEPRECATED & DEFUNCT

    o   .Alias has been removed from all R sources and deprecated.

    o   reshapeLong(), reshapeWide() are deprecated in favour of reshape().

    o   Previously deprecated functions read.table.url(), scan.url(),
    source.url(), httpclient() and parse.dcf() are defunct.
    Method "socket" for download.file() no longer exists.


DOCUMENTATION CHANGES

    o   `Writing R Extensions' has a new chapter on generic/method
    functions.


UTILITIES

    o   New package `tools' for package development and administration
    tools, containing the QC tools checkFF(), codoc() and undoc()
    previously in package base, as well as the following new ones:

    -  checkAssignFuns() for checking whether the final argument of
       assignment functions in a package is named `value'.

    -  checkDocArgs() for checking whether all arguments shown in
       \usage of Rd files are documented in the corresponding
       \arguments.

    -  checkMethods() for checking whether all methods defined in a
       package have all arguments of their generic.

    -  checkTnF() for finding expressions containing the symbols `T'
       and `F'.

    o   R CMD Rd2dvi has more convenient defaults for its output file.

    o   R CMD check now also fully checks the Depends field in the
    package DESCRIPTION file.  It also tests for syntax errors in
    the R code, whether all methods in the code have all arguments
    of the corresponding generic, for arguments shown in \usage but
    not documented in \arguments, and whether assignment functions
    have their final argument named `value'.


C-LEVEL FACILITIES

    o   arraySubscript and vectorSubscript are now available to package
    users.  All "array-like" packages can use a standard method for
    calculating subscripts.

    o   The C routine type2symbol, similar to type2str, returns a symbol
    corresponding to the type supplied as an argument.

    o   The macro SHLIB_EXT now includes `.', e.g. ".so" or ".dll",
    since the Mac uses "Lib" without a `.'.

    o   New Fortran entry points rwarn() and rexit() for warnings and
    error exits from compiled Fortran code.

    o   A new serialization mechanism is available that can be used to
    serialize R objects to connections or to strings.  This
    mechanism is used for the version 2 save format.  For now,
    only an internal C interface is available.

    o   R_tryEval() added for evaluating expressions from C code with
    errors handled but guaranteed to return to the calling C
    routine. This is used in embedding R in other applications and
    languages.

    o   Support for attach()'ing user-defined tables of variables
    is available and accessed via the RObjectTables package
    currently at http://www.omegahat.org/RObjectTables.


BUG FIXES

    o   Fixed share/perl/massage-examples.pl to detect instances of
    par() at the very start of a line.

    o   Fixed Pearson residuals for glms with non-canonical
    link.(PR#1123). Fixed them again for weights (PR#1175).

    o   Fixed an inconsistency in the evaluation context for on.exit
    expressions between explicit calls to `return' and falling off
    the end returns.

    o   The code in model.matrix.default() handling contrasts was
    assuming a response was present, and so without a response was
    failing to record the contrasts for the first variable if it
    was a factor.

    o   diffinv() could get the time base wrong in some cases.

    o   file.append() was opening all files in text mode: mattered
    on Windows and classic Macintosh. (PR#1085)

    o   f[] <- g  now works for factor f.

    o   substr<-() was misbehaving if the replacement was too short.

    o   The version of `packages.html' generated when building R or
    installing packages had an incorrect link to the style sheet.
    The version used by help.start() was correct. (PR#1090)

    o   rowsum() now gives character (not factor codes) as rownames.
    (PR#1092)

    o   plot.POSIX[cl]t now respect the `xaxt' parameter.

    o   It is now possible to predict from an intercept-only model:
    previously model.matrix.default() objected to a 0-column
    model frame.

    o   c.POSIXct was not setting the right classes in 1.3.x.

    o   cor(*, use = "all.obs") <= 1 is now guaranteed which ensures
    that sqrt(1 - r^2) is always ok in cor.test(). (PR#1099)

    o   anova.glm() had a missing drop=FALSE and so failed for some
    intercept-less models.

    o   predict.arima0() now accepts vector as well as matrix
    `newxreg' arguments.

    o   cbind(d,f) now works for 0-column dataframes. This fixes PR#1102.

    o   plot(ts(0:99), log = "y") now works

    o   method "gnudoit" of bug.report() was incorrectly documented as
    "gnuclient" (PR#1108)

    o   saving with ascii=TRUE mangled backslashes. (PR#1115)

    o   frac(,) {and others} now adds a gap appropriately. (PR#1101)

    o   logLik.lm() now uses the correct "df" (nlme legacy code).

    o   closeAllConnections() works again, and closes all sink() diversions.

    o   sink(type="message") works again.

    o   sink.number was (accidentally) returning the result invisibly.

    o   as.POSIXct("NA") (or ..lt) now work; hence, merge(*, all=TRUE) now
    works with dataframes containing POSIXt date columns.

    o   integer(2^30+1) and similar ones do not segfault anymore but duly
    report allocation errors.

    o   seq(0, 0, 1) now works (PR#1133).

    o   reshapeWide() got it wrong if the "i" factor was not sorted (the
    function is now deprecated since reshape() is there, but the bug
    still needed fixing...)

    o   PR#757 was fixed incorrectly, causing improper subsetting of
    pch etc. in plot.formula().

    o   library() no longer removes environments of functions that are
    not defined in the top-level package scope.  Also, packages
    loaded by require() when sourcing package code are now visible
    in the remaining source evaluations.

    o   names(d) <- v  now works (again) for "dist" objects d. (PR#1129)

    o   Workarounds for problems with incompletely specified date-times
    in strptime() which were seen only on glibc-based systems
    (PR#1155).

    o   promax() was returning the wrong rotation matrix. (PR#1146)

    o   The [pqr]signrank and [pqr]wilcox functions failed to check that
    memory has been allocated (PR#1149), and had (often large)
    memory leaks if interrupted.  They now can be interrupted on
    Windows and Mac OS and don't leak memory.

    o   range(numeric(0)) is now c(NA, NA) not NA.

    o   round(x, digits) for digits <= 0 always gives an integral
    answer.  Previously it might not due to rounding errors in
    fround. (PR#1138/9)

    o   Several memory leaks on interrupting functions have been
    circumvented.  Functions lqs() and mve() can now be
    interrupted on Windows and Mac OS.

    o   image() was finding incorrect breakpoints from
    irregularly-spaced midpoints. (PR#1160)

    o   Use fuzz in the 2-sample Kolmogorov-Smirnov test in package
    ctest to avoid rounding errors (PR#1004, follow-up).

    o   Use exact Hodges-Lehmann estimators for the Wilcoxon tests in
    package ctest (PR#1150).

    o   Arithmetic which coerced types could lose the class
    information, for example `table - real' had a class attribute
    but was not treated as a classed object.

    o   Internal ftp client could crash R under error conditions such
    as failing to parse the URL.

    o   Internal clipping code for circles could attempt to allocate
    a vector of length -1 (related to PR#1174)

    o   The hash function used internally in match(), unique() and
    duplicated() was very inefficient for integers stored as
    numeric, on little-endian chips.  It was failing to hash the
    imaginary part of complex numbers.

    o   fifo() no longer tries to truncate on opening in modes
    including "w". (Caused the fifo example to fail on HP-UX.)

    o   Output over 1024 characters was discarded from the GNOME
    console.

    o   rug() now correctly warns about clipped values also for logarithmic
    axes and has a `quiet' argument for suppressing these (PR#1188).

    o   model.matrix.default was not handling correctly contrasts.arg
    which did not supply a full set of contrasts (PR#1187).

    o   The `width' argument of density() was only compatible with S
    for a Gaussian kernel: now it is compatible in all cases.

    o   The rbinom() C code had a transcription error from the original
    Fortran which led to a small deviation from the intended
    distribution. (PR#1190)

    o   pt(t, , ncp=0) was wrong if t was +/-Inf.

    o   Subsetting grouping factors gave incorrect degrees of freedom
    for some tests in package ctests. (PR#1124)

    o   writeBin() had a memory leak.

    o   qbeta(0.25, 0.143891, 0.05) was (incorrectly) 3e-308. (PR#1201)

    o   Fixed alignment problem in ppr.f on Irix. (PR#1002, 1026)

    o   glm() failed on null binomial models. (PR#1216)

    o   La.svd() with nu = 0 or nv = 0 could fail as the matrix passed
    to DGESVD was not of dimension at least one (it was a vector).

    o   Rownames in xcoef and ycoef of cancor() were wrong if x or y
    was rank-deficient.

    o   lqs() could give warnings if there was an exact fit. (PR#1184)

    o   aov() didn't find free-floating variables for Error() terms when
    called from inside another function

    o   write.table() failed if asked to quote a numerical matrix with
    no row names. (PR#1219)

    o   rlnorm( *, *, sd=0) now returns the mean,
    rnbinom(*, *, prob=1) gives 0, (PR#1218).


    **************************************************
    *                        *
    *          1.3 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 1.3.1


NEW FEATURES

    o   massage-examples is now a Perl script and about 50x faster.

    o   On Unix(-alike) systems the default pager is now determined
    during configuration, and is `less' if available, otherwise
    `more' (and not `more -s' as previously).

    o   configure now tests for strptime functions that fail on inputs
    before 1970 (found on Irix).  It no longer checks for the SCSL
    and SGIMATH libraries on Irix.

    o   New formula interface to cor.test() in package ctest.

    o   "NA" is now a valid color name (as NA has been a valid integer
    color).

    o   pairs() function has a new `gap' argument for adjusting the
    spacing between panels.

    o   R CMD check has a new test for unbalanced braces in Rd files.

    o   readBin() has a new argument `signed' to simplify reading
    unsigned 8- and 16-bit integers.

    o   New capabilities() option "cledit".

    o   Modified restore code to give clearer error messages in some cases.



BUG FIXES

    o   Fixed placement of mtext() output (including tick labels)
    when text is perpendicular to axis AND axis is logged
    (PR#997 and PR#865).

    o   rect() and arrows() now respond to setting of par(xpd)

    o   abline() now responds to the setting of par(xpd), including
    via the "..." argument (PR#750).

    o   Using cat() to a gzfile() connection no longer compresses poorly
    (at the expense of some latency).

    o   The `exact' p-values for the two-sided two-sample
    Kolmogorov-Smirnov test in 1.3.0 were for the wrong tail.

    o   In the HTML conversion of .Rd files, \link[foo]{bar} was
    ignoring [foo] outside \code{} statements.

    o   lm.influence(), plot.lm(), influence.measures() and the
    related diagnostic measures now handle lm() fits with
    na.action=na.exclude.

    o   Eliminated segmentation fault in while loops without braces
    in their bodies.

    o   barplot did not accept NA heights unless ylim was specified.

    o   predict.lm() failed with single-column model matrix (PR#1018).

    o   legend() now also works with 100s of entries (thanks to
    M. Schlather).

    o   A long-standing bug leading to inaccuracy in polyroot() has been
    fixed (PR#751)

    o   A very old bug with eval() not handling a numeric `envir' argument
    according to specifications has been fixed. [This in particular
    broke ported S-PLUS code that used eval(..., sys.parent())]. Also,
    eval() now checks that a numeric `envir' has length one to prevent
    accidental passing of matrices instead of data frames.

    o   The C code underlying nlm(f, *) now also works correctly when f(x)
    returns +Inf {nlm() itself *did* work!}.

    o   pdfcolor.tex is now included in doc/manual, as pdftex seems no
    longer to include it.

    o   Fixed protect bug in save() code that might have been responsible
    for occasionally saving corrupted workspaces.

    o   capabilities("X11") now reports if X11 is available in this
    session, thereby giving the correct answer if --gui=none was
    specified.

    o   Rd files with ,, inside \code{} were treating this as a
    ligature in LaTeX: now corrected in Rdconv.

    o   dlopen on Compaq Tru64 was finding the wrong entry points:
    worked-around by using different names in the modules.

    o   plot.mts() allows a type argument again (broken in 1.3.0: PR#1010).

    o   scan() has a limit of 8190 on the length of char strings, and
    now warns if it is exceeded (when the string has always been
    truncated).

    o   par adj now accepts values of integer mode (such as text(adj=0:1)).

    o   horizontal adj values outside the range [0, 1] are now
    supported on devices with canHAdj==1 (Windows devices, xfig).

    o   xtabs() without a response was ignoring drop.unused.levels=TRUE.

    o   readLines(ok=FALSE) was not generating an error when it should.

    o   princomp(covmat=) has been broken.

    o   Many documentation clean-ups: formerly undocumented arguments,
    use of T/F rather than TRUE/FALSE.

    o   df[] <- foo now works for data frames (PR#1047).

    o   nargs() was documented incorrectly.

    o   Using seq.POSIXt(by="2 weeks") was stepping by single weeks
    (PR#1046).

    o   dummy.coef.lm was not handling complex terms like z:I(x).
    It now either does or warns and gives NAs (PR#1048).

    o   predict.mlm() was broken (PR#1049).

    o   ksmooth (in package modreg) was using points to the left
    of the kernel under some circumstances (PR#1050).

    o   attr(x, "class") <- character(0) was segfaulting.  Similar problem
    with "comment" attribute also fixed.

    o   loadings() results from PCA (as well as from factor analysis)
    are now printed by print.loadings.

    o   Using chol() on an non-positive-definite matrix gave an
    misleading error message (PR#1061).

    o   as.character() on a language object was truncating to about 70
    characters: the internal limit (from the deparse code) is now the
    maximum possible 500.

    o   X11() device was failing to re-initialize correctly after all
    devices were shut down, so the colortype could not be changed
    and some systems showed protocol errors. (PR#1065)

    o   Converting Inf and NaN values by as.character() and reading by
    scan() was supported by strtod on some platforms and not
    others: these are now handled by R itself. (PR#1072)

    o   hclust(dist(1)) doesn't loop infinitely but signal an error.

    o   cutree() can now cut outside heights and does not return
    garbage for h >= max(tree$heights) anymore. (PR#1067)

    o   interaction(...,drop=F) returned an object looking like a
    factor, but with storage mode "double", confusing
    model.matrix.default (PR#1003)

    o   splineDesign(*, ord=) and others give better error messages for
    wrong `ord' (and have slightly improved documentation).


        CHANGES IN R VERSION 1.3.0


NEW FEATURES

    o   Changes to connections:

    -  New function url() to read from URLs.  file() will also
       accept URL specifications, as will all the functions which
       use it.

    -  file connections can now be opened for both reading and writing.

    -  Anonymous file connections (via file()) are now supported.

    -  New function gzfile() to read from / write to compressed files.

    -  New function fifo() for connections to / from fifos (on Unix).

    -  Text input from file, pipe, fifo, gzfile and url connections
       can be read with a user-specified encoding.

    -  New functions readChar() and writeChar() to read character
       strings with known lengths and no terminators, and to write
       user-specified lengths from strings.

    -  sink() now has a stack of output connections, following S4.

    -  sink() can also be applied to the message stream, to capture
       error messages to a connection.  Use carefully!

    -  seek() has a new `origin' argument.

    -  New function truncate() to truncate a connection open for
       writing at the current position.

    -  Socket connections via function socketConnection().

    -  The `blocking' argument for file, fifo  and socket connections
       is now operational.

    o   Changes to date/time classes and functions:

    -  Date/time objects now all inherit from class "POSIXt".

    -  New function difftime() and corresponding class for date/time
       differences, and a round() method.

    -  Subtraction and logical comparison of objects from different
       date/time classes is now supported.  NB: the format for the
       difference of two objects of the same date/time class has
       changed, but only for objects generated by this version, not
       those generated by earlier ones.

    -  Methods for cut(), seq(), round() and trunc() for date/time
       classes.

    -  Convenience generic functions weekdays(), months(), quarters()
       and julian() with methods for "POSIXt" objects.

    o   Coercion from real to integer now gives NA for out-of-range
    values, rather than the most extreme integer of the same sign.

    o   The Ansari-Bradley, Bartlett, Fligner-Killeen, Friedman,
    Kruskal-Wallis, Mood, Quade, t, and Wilcoxon tests as well as
    var.test() in package ctest now have formula interfaces.

    o   Matrix multiplication functions %*% and crossprod() now use a
    level-3 BLAS routine dgemm.  When R is linked with the ATLAS
    or other enhanced BLAS libraries this can be substantially
    faster than the previous code.

    o   New functions La.eigen() and La.svd() for eigenvector and
    singular value decompositions, based on LAPACK.  These are
    preferred to eigen() and svd() for new projects and can make
    use of enhanced BLAS routines such as ATLAS.  They are
    used in cancor(), cmdscale(), factanal() and princomp()
    and this may lead to sign reversals in some of the output of
    those functions.

    o   Provided the Fortran compiler can handle COMPLEX*16, the
    following routines now handle complex arguments, based on
    LAPACK code.

    qr, qr.coef, qr.solve, qr.qy, qr.qty, solve.default, svd, La.svd.



    o   aperm() uses strides in the internal C code and so is
    substantially faster (by Jonathan Rougier).

    o   The four bessel[IJKY](x,nu) functions are now defined for nu < 0.

    o   [dpqr]nbinom also accept an alternative parametrization via the
    mean and the dispersion parameter (thanks to Ben Bolker).

    o   Generalised "birthday paradox" functions [pq]birthday.

    o   boxplot() and bxp() have a new argument `at'

    o   New function capabilities() to report optional capabilities
    such as jpeg, png, tcltk, gzfile and url support.

    o   New function checkFF() for checking foreign function calls.

    o   New function col2rgb() for color conversion of names, hex, or
    integer.

    o   coplot() has a new argument `bar.bg' (color of conditioning bars),
    gives nicer plots when the conditioners are factors, and allows
    factors for x and y (treated almost as if unclass()ed) using new
    argument `axlabels'.  [original ideas by Thomas Baummann]

    o   `hessian' argument added to deriv() and its methods.  A new function
    deriv3() provides identical capabilities to deriv() except that
    `hessian' is TRUE by default.
    deriv(*, *, func = TRUE) for convenience.

    o   New dev.interactive() function, useful for setting defaults for
    par(ask=*) in multifigure plots.

    o   dist() in package mva can now handle missing values, and
    zeroes in the Canberra distance.

    o   The default method for download.file() (and functions which use
    it such as update.packages()) is now "internal", and uses code
    compiled into R.

    o   eigen() tests for symmetry with a numerical tolerance.

    o   New function formatDL() for formatting description lists.

    o   New argument `nsmall' to format.default(), for S-PLUS
    compatibility (and used in various packages).

    o   ?/help() now advertises help.search() if it fails to find a topic.

    o   image() is now a generic function.

    o   New function integrate() with S-compatible call.

    o   New function is.unsorted() the C version of which also speeds up
    .Internal(sort()) for sorted input.

    o   is.loaded() accepts an argument PACKAGE to search within
    a specific DLL/shared library.

    o   Exact p-values are available for the two-sided two-sample
    Kolmogorov-Smirnov test.

    o   lm() now passes `...' to the low level functions for regression
    fitting.

    o   Generic functions logLik() and AIC() moved from packages nls
    and nlme to base, as well as their *.lm methods.

    o   New components in .Machine give the sizes of long, long long and
    long double C types (or 0 if they do not exist).

    o   merge.data.frame() has new arguments, `all[.xy]' and `suffixes',
    for S compatibility.

    o   model.frame() now calls na.action with the terms attribute set
    on the data frame (needed to distiguish the response, for
    example).

    o   New generic functions naresid(), napredict() and naprint()
    (formerly in packages survival5 and MASS, also used in rpart).
    Also na.exclude(), a variant on na.omit() that is handled
    differently by naresid() and napredict().

    The default, lm and glm methods for fitted, residuals, predict
    and weights make use of these.

    o   New function oneway.test() in package ctest for testing for
    equal means in a one-way layout, assuming normality but not
    necessarily equal variances.

    o   options(error) accepts a function, as an alternative to
    an expression.  (The Blue Book only allows a function; current
    S-PLUS a function or an expression.)

    o   outer() has a speed-up in the default case of a matrix outer
    product (by Jonathan Rougier).

    o   package.skeleton() helps with creating new packages.

    o   New pdf() graphics driver.

    o   persp() is now a generic function.

    o   plot.acf() makes better use of white space for `nser > 2', has
    new optional arguments and uses a much better layout when more
    than one page of plots is produced.

    o   plot.mts() has a new argument `panel' providing the same
    functionality as in coplot().

    o   postscript() allows user-specified encoding, with encoding files
    supplied for Windows, Mac, Unicode and various others, and with
    an appropriate platform-specific default.

    o   print.htest() can now handle test names that are longer than
    one line.

    o   prompt() improved for data sets, particularly non-dataframes.

    o   qqnorm() is now a generic function.

    o   read.fwf() has a new argument `n' for specifying the number of
    records (lines) read in.

    o   read.table() now uses a single pass through the dataset.

    o   rep() now handles lists (as generic vectors).

    o   scan() has a new argument `multi.line' for S compatibility,
    but the default remains the opposite of S (records can cross
    line boundaries by default).

    o   sort(x) now produces an error when x is not atomic instead of
    just returning x.

    o   split() now allows splitting on a list of factors in which case
    their interaction defines the grouping.

    o   stl() has more optional arguments for fine tuning, a summary()
    and an improved plot() method.

    o   New function strwrap() for formatting character strings into
    paragraphs.

    o   New replacement functions substr<-() and substring<-().

    o   Dataset swiss now has row names.

    o   Arguments `pkg' and `lib' of system.file() have been renamed to
    `package' and `lib.loc', respectively, to be consistent with
    related functions.  The old names are deprecated.  Argument
    `package' must now specify a single package.

    o   The Wilcoxon and Ansari-Bradley tests now return point
    estimators of the location or scale parameter of interest along
    with confidence intervals for these.

    o   New function write.dcf() for writing data in Debian Control File
    format.  parse.dcf() has been replaced by (much faster) internal
    read.dcf().

    o   Contingency tables created by xtabs() or table() now have a
    summary() method.

    o   Functions httpclient(), read.table.url(), scan.url() and
    source.url() are now deprecated, and hence method="socket' in
    download.file() is.  Use url connections instead: in
    particular URLs can be specified for read.table(), scan() and
    source().

    o   Formerly deprecated function getenv() is now defunct.

    o   Support for package-specific demo scripts (R code).  demo() now
    has new arguments to specify the location of demos and to allow
    for running base demos as part of `make check'.

    o   If not explicitly given a library tree to install to or remove
    from, respectively, R CMD INSTALL and R CMD REMOVE now operate
    on the first directory given in `R_LIBS' if this is set and
    non-null, and the default library otherwise.

    o   R CMD INSTALL and package.description() fix some common problems
    of DESCRIPTION files (blank lines, ...)

    o   The INSTALL command for package installation allows a `--save'
    option.  Using it causes a binary image of the package contents
    to be created at install time and loaded when the package is
    attached.  This saves time, but also uses a more standard way
    of source-ing the package.  Packages that do more than just
    assign object definitions may need to install with `--save'.
    Putting a file `install.R' in the package directory makes
    `--save' the default behavior.  If that file is not empty, its
    contents should be R commands executed at the end of loading
    the image.

    There is also a new command line option `--configure-vals' for
    passing variables to the configure script of a package.

    o   R CMD check now also checks the keyword entries against the list
    of standard keywords, for code/documentation mismatches (this
    can be turned off by the command line option `--no-codoc'), and
    for sufficient file permissions (Unix only).  There is a new check
    for the correct usage of library.dynam.

    It also has a new command line option `--use-gct' to use
    `gctorture(TRUE)' when running R code.

    o   R CMD Rd2dvi has better support for producing reference manuals
    for packages and package bundles.

    o   configure now tests for the versions of jpeg (>= 6b), libpng (>=
    1.0.5) and zlib (>= 1.1.3).  It no longer checks for the
    CXML/DXML BLAS libraries on Alphas.

    o   Perl scripts now use Cwd::cwd() in place of Cwd::getcwd(), as
    cwd() can be much faster.

    o   R::Dcf.pm can now also handle files with more than one record
    and checks (a little bit) for continuation lines without leading
    whitespace.

    o   New manual `R Installation and Administration' with fuller
    details on the installation process: file `INSTALL' is now a
    brief introduction referencing that manual.

    o   New keyword `internal' which can be used to hide objects that
    are not part of the API from indices like the alphabetical lists
    in the HTML help system.

    o   Under Unix, shlib modules for add-on packages are now linked
    against R as a shared library (`libR') if this exists.  (This
    allows for improved embedding of R into other applications.)

    o   New mechanism for explicitly registering native routines in a
    DLL/shared library accessible via .C(), .Call(), .Fortran() and
    .External().  This is potentially more robust than the existing
    dynamic lookup, since it checks the number of arguments, type of
    the routine.

    o   New mechanism allowing registration of C routines for converting
    R objects to C pointers in .C() calls.  Useful for references to
    data in other languages and libraries (e.g. C and hdf5).

    o   The internal ftp/http access code maintains the event loop, so
    you can download whilst running tcltk or Rggobi, say.  It can
    be hooked into package XML too.



BUG FIXES

    o   boxplot.stats() now returns the correct `n' component in the case of
    NAs. This also affects `conf', the notch length (PR#967).

    o   the "coef = 0" fix (PR#917) to boxplot.stats() broke
    boxplot.stats(do.out = FALSE).

    o   curve(* , add = FALSE) now works correctly again if used when
    par("xlog") is true.

    o   Printing a dataframe with a column called "row.names" or
    "check.names" omitted the column.

    o   data.frame(a=1, b=2)[1, -(1:2)]  now works, giving an empty
    data frame.  as.matrix(df), deparse(), dput() and str() now work
    for empty data frames.

    o   dbeta() could return zero erroneously on some platforms where
    roundoff makes e.g. 1.3 - 1 > (1.3 + 1) - 2 (Solaris and HP-UX
    were affected). (PR#970)

    o   pointer protection in deriv() (PR#953)

    o   expand.data.frame() didn't interpret its subset= argument in
    the data frame. (PR#979 et al., fix from Greg Warnes)

    o   format() on character arrays was losing the dimensions.

    o   hist.default() was ignoring its `probability' argument for
    equi-spaced breaks.

    o   Occasionally insane results of is.nan(NULL), is.na(NULL) (PR#952)

    o   ks.test() in package ctest now had the one-sided alternatives in
    the one-sample case reversed (PR#934).

    o   ls.str() and lsf.str() now work better with pos > 1, and are
    correctly documented (PR#899).

    o   plot(1:2, xaxs = "i") now gives both "1.0" and "2.0" x-axis labels.

    o   promax() was not using its `m' argument (m=4 was always used).

    o   qr.X() was ignoring pivoting (which happens only in
    rank-deficient cases).

    o   sign of Pearson residuals was wrong for reciprocal link (PR#862)

    o   The algorithm for determining starting estimates in the SSfpl
    self-starting model in the nls package was changed following a
    suggestion by Phillipe Grosjean <phgrosje@ulb.ac.be>

    o   svd(x)$v was not a matrix when x had only 1 row (PR#963).

    o   symnum(x) behaves more reasonably for zero length or non-numeric x.

    o   The implementation of sys.function() did not work when the
    current function was already a function (e.g., a closure).

    o   tapply() had "..." argument after simplify=, which made it
    incompatible with S when passing additional unnamed arguments to FUN.

    o   use formula environments in xy.coords() (PR#932)


    o   Unused entry points in `src/appl/fortran.c' and `Fortran.h'
    have been removed.

    o   Unused directory `src/f2clib' and header `src/include/R_ext/f2c.h'
    have been removed.

    o   The documentation and parts of the parse/deparse mechanism
    indicated that there was an operator `%'. Corrected (PR#925).

    o   Fixed some cases where the deparser gave code that threw a
    syntax error when parsed.

    o   Using a pushback with more than two lines at a time could crash.

    o   Eliminate or warn on use of .Last.value in help examples, since
    .Last.value is only set from top-level expressions (and was
    incorrectly documented).

    o   asInteger was ignoring integer overflow (PR#929 and others).

    o   approxfun(*, ties = <fun>) could return garbage when there were ties;
    ties = "ordered" was incorrect--also for approx()-- when there was
    a tie at the very right.


    **************************************************
    *                        *
    *          1.2 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 1.2.3

NEW FEATURES

    o   Support for configuration and building the Unix version of R
    under Mac OS X.  (The `classic' Macintosh port is `Carbonized'
    and also runs under that OS.)

    o   dotchart() and stripchart() become the preferred names for
    dotplot() and stripplot(), respectively.  The old names are now
    deprecated.

    o   Functions in package ctest now consistently use +/-Inf rather
    than NA for one-sided confidence intervals.



BUG FIXES

    o   `Writing R Extensions' incorrectly described an entry point
    `fmod' which was not included in the R binary on most systems.
    The entry point has been removed, to avoid any confusion with
    the standard C entry point of the same name.

    o   Printing of kernels in package ts was unreliable: the internal
    representation was correct.

    o   A problem with "..." in substitutions has been fixed (PR#860).

    o   Various strangeness with match.call() and "..." has been fixed
    (PR#813).

    o   sys.function() could return an unevaluated promise, messing up
    formals() and match.arg() (PR#872)

    o   Deparsing or dumping attributes with non-standard names was
    not quoting the name (PR#871).

    o   We now use match.fun() in tapply() so that it won't get trapped
    by non-functions masking FUN.

    o   The `nmax' argument in scan(what=list(...), nmax=300) was
    counting blank lines, although they were said to be ignored.
    This affected using read.table on files with blank lines in
    1.2.2, and using `n' with `what' a list.

    o   as.numeric(), as.integer(), as.complex() returned zero on
    completely blank strings. Now they return NA. (PR#870)

    o   Overflow in deparse() in terms(formula) with very long left-hand
    side fixed.  (PR#873)

    o   lowess(c(1:3, 7:5)) now should give the same on different platforms,
    since now in the C code (int)(f*n + 1e-7) is used.

    o   curve(*, add = TRUE) now works properly with active log scales.

    o   rt() could give different results on different platforms as
    the order of evaluation of calls to random numbers was not
    fully defined.  This affected the Mac port.

    o   ppr() inside nested calls sometimes failed to find objects
    due to a typo in the eval call in ppr.formula.

    o   qchisq(0.025, 31, ncp=1, lower.tail=FALSE) now works.  (PR#875)

    o   dchisq with integer x, df and ncp sometimes gave incorrect
    results on i686 Linux.

    o   Cancelling a quit caused quit() not to ask next time it was called.

    o   Some complicated operations with "..." didn't work.

    o   Missingness was not correctly propagated in some cases.

    o   eigen() segfaulted on 0-dimensional matrices.

    o   nls( ~ rhs, ..)  now works (formula with empty left hand side).

    o   The fuzz added in hist() for 1.2.2 broke the logic testing for
    equidistant breakpoints.

    o   Calls to replacement functions ("f<-") lost argument names
    (PR#888).

    o   is.na() and is.nan() gave random results on lists containing
    zero-length objects.

    o   cor(), cov() and var() gave a result on a zero-length argument
    (now an error).  cov(x, NULL) and cor(x, NULL) are now errors
    if x is a vector.  (PR#883).

    o   ?smooth.spline now properly describes `spar' which is *not*
    lambda.  smooth.spline(1:20, spar = 50) gives an error instead
    of silent nonsense.  print.smooth.spline() now makes use of a
    digits argument.

    o   Confidence intervals for wilcox.test() had the samples reversed
    in the two-sided case (PR#895), and sometimes got continuity
    correction wrong (PR#896).

    o   Using out-of-range font values in text() on a postscript()
    device on Windows could crash.  .ps.prolog was incorrectly
    named in some of the documentation.  (PR#914)

    o   Warning messages of > 8191 chars could crash R (PR#902), as
    could errors.  Now they are truncated to 8191 chars on
    machines which have vsnprintf (almost all).

    o   range() now works properly for date-time objects.

    o   contour() could loop infinitely (PR#897).

    o   R-lang manual had precedence of %% wrong (PR#879).

    o   try() constructs lost protection of R_alloc'ed memory (PR#877).

    o   Documented that as.numeric() dispatches to as.double.foo methods
    (PR#918.1).

    o   httpclient() (and the "socket" method of download.file) skipped ^V
    rather than ^Z in input (if drop.ctrl.z = TRUE, the default).

    o   boxplot(*, range = 0) and boxplot.stats(*, coef = 0) now
    don't return outliers anymore (PR#917).

    o   segmentation fault with tmp[[1,]] (PR#904)

    o   incorrect "..." handling in plot.factor (PR#830)



        CHANGES IN R VERSION 1.2.2


NEW FEATURES

    o   The Macintosh port becomes a full member of the R family and its
    sources are incorporated as from this release.
    See `src/macintosh/INSTALL' for how that port is built.

    o   The API header files and export files `R.exp' are released under
    LGPL rather than GPL to allow dynamically loaded code to be
    distributed under licences other than GPL.

    o   postscript() and xfig() devices now make use of genuine Adobe
    afm files, and warn if characters are used in string width or
    height calculations that are not in the afm files.

    o   Configure now uses a much expanded search list for finding a
    FORTRAN 77 compiler, and no longer disallows wrapper scripts for
    this compiler.

    o   New Rd markup \method{GENERIC}{CLASS} for indicating the usage
    of methods.

    o   print.ftable() and write.ftable() now have a `digits' argument.

    o   undoc() has a new `lib.loc' argument, and its first argument is
    now called `package'.


BUG FIXES

    o   The behaviour of polygon() with NA coordinate values is now
    documented.

    o   polygon() now correctly recycles the values of border and lty.
    Thanks to Denis White.

    o   readBin() and writeBin() were not always making good use of
    long types (such as long double) when available.

    o   The C function fprec was not handling very small denormalized
    numbers correctly.  This affected the way they were printed on
    Windows, for example.

    o   legend() now draws points after lines allowing "empty dots on
    lines" thanks to Ben Bolker.

    o   Setting row/column names preserves the names of the dimnames
    (PR#809).

    o   Using help() after help.start() on Unix was failing to find
    the linked help files if more than one package was loaded (as
    now happens by default).

    o   scan now sets the maximal sizes internally when `what' is a list
    from `nmax' and failing that `nlines'.  This uses memory much
    more efficiently if `nmax' is specified.  read.table() makes use
    of this.

    o   The FORTRAN code used is now much closer to ANSI compliance.

    o   cov.rob() gives a useful error message on nearly-degenerate
    matrices.

    o   summary() on a data frame was computing results to a precision
    based on options("digits") (default 4), not its `digits'
    argument.  Now results are computed to full precision but
    formatted using `digits'.

    o   summary(m, digits=*) also properly works for matrices.

    o   When returning parameters from .C/.Fortran, the OBJECT field was
    not copied even though class attributes were (PR#834).

    o   There was a spurious warning from the X11 driver on 100 dpi
    displays (PR#840, fix from Tom Vogels).

    o   scan() was reading all-blank numeric fields as zero.
    Should be read as NA, and are now.

    o   dnbinom(*, size, *) now works again for size < 1 (PR#842).

    o   dgeom(*, p = 0) and pgeom(*, p = 0) don't give NaN anymore.

    o   smooth.spline() allows zero weights again.

    o   aperm() fix for list arg; aperm(*, resize=FALSE) now works with
    dimnames (by dropping them).

    o   The xfig() driver was not closing unclosed polygons as the R
    driver expected (but did not document).

    o   spec.ar() now handles order 1 correctly.

    o   Add fuzz to prevent anomalies with hist().

    o   print.matrix() checks length of rowlab, collab (PR#850)

    o   add1.glm() was using the wrong weights for binomial glms in
    cbind(successes, failures) ~ lhs  form, since 1.2.0.

    o   Eliminate packing of the `structRstart' structure, which was
    probably non-ISO C and caused some compilers problems (PR#852).

    o   all.vars() no longer returns empty `names' (PR#829).

    o   dotplot() passes `cex' down to the mtext() calls used for the
    labels (PR#816).

    o   The pictex() driver was drawing polylines incorrectly (PR#845).

    o   In the event of a fatal X11 error (this can be generated under FVWM
    by selecting Destroy), the input handler would loop, leaving the
    "please save work and shut down R" advice useless. There should now
    be an opportunity for a relatively graceful exit.

    o   polyroot() misbehaved if all coefficients except that of the highest
    degree term were zero (partial fix for PR#751)

    o   spline(method="natural") was extrapolating incorrectly (both
    to the left and to the right, for different reasons).  (PR#653)

    o   Line clipping was incorrect for log-plots. (PR#839)

    o   Add explicit test for rank-deficient residuals to summary.manova().

    o   The help search data bases did not contain the data from the
    platform-specific help files.

    o   parse.dcf() and hence package.contents() now work with a warning
    for empty contents' files.

    o   which.min() and which.max() now properly work for 0-length and
    NA-only arguments.



        CHANGES IN R VERSION 1.2.1


NEW FEATURES

    o   New functions factanal(), varimax(), promax() and examples in
    package mva.

    o   New functions readBin() and writeBin() to transfer binary data
    to and from connections.

    o   merge() is partially moved to C to reduce its memory usage.

    o   library(help = PKG) now displays the contents of the package's
    DESCRIPTION file in addition to its INDEX.

    o   Sd2Rd can handle S4-style documentation too: see `Writing R
    Extensions'.

    o   prompt() now also works with a character argument (useful for
    producing many *.Rd files in a loop).

    o   The Unix front-end shell script now ignores a value for R_HOME
    found in the environment.

    o   Connections functions such as file() now accept a description of
    length > 1, with a warning.

    o   All text-mode connections now accept input with LF, CR or CRLF
    line endings.  This means that readLines() can be used on DOS
    files and source() on Mac files, for example.

    Also, CRLF-terminated files can be used as stdin on Unix, and
    files with last lines without an EOL mark can be used as stdin
    and source()-ed on Unix and Windows.

    o   DESCRIPTION file has a new recommended `Maintainer:' field.

    o   stars() now uses a larger "cex" for the labels, and cex and lwd
    are now arguments. Further, the argument names (xlim, ylim,
    axes) are now consistent with other plot functions.  The key symbol
    is not clipped anymore into the plot region by default.

    o   Date-time quantities are now printed with the timezone, if known.

    o   R CMD build now ignores all files specified (via Perl regexps)
    in file `.Rbuildignore' in the top-level source directory of a
    package.

    o   Horizontal boxplots are possible with  horizontal = TRUE.

    o   all.equal() on lists now compares them as generic vectors, that
    is they are equal if have identical names attributes and
    all components are equal.

    o   Invalid lines in .Renviron now give warnings when R is started.

    o   Argument `na.last' implemented for rank().


BUG FIXES

    o   There have been many small improvements/corrections in the
    documentation. In particular:

    The help pages for lm and glm now describe the fitted objects.

    is.vector() is now correctly documented: complex vectors and
    lists return TRUE.

    o   The default sortedXyData constructor in the nls package now
    handles missing values in the response.

    o   On startup, file .RData is now restored *after* the site and
    user profiles are loaded (if at all).  This is as previously
    documented.

    o   as.POSIXlt (and *ct) operating on a character string
    only extracted the date (and not the time) in some cases.

    o   as.POSIXct() on character objects was ignoring `tz'.

    o   codoc(ignore.generic.functions = TRUE) works again.

    o   Explicitly close files opened by sink() for OSes (such as
    Windows) that benefit from it.

    o   Prevent closing the sink connection (which would be a silly
    thing to do).

    o   showConnections(all = TRUE) was sometimes counting connections
    incorrectly, and so not showing some closed ones.

    o   ts(1:10, start= c(1999,6), end = c(2000,3), frequency = 12)
    now prints both years (as it should).

    o   Monthly multivariate time series now print proper month names
    in all cases.

    o   print.stl(), print.princomp() and print.prcomp() now pass on
    `digits =' etc.

    o   prompt() now produces a \usage{} line for data frames and
    other datasets, as documented in `Writing R Extensions'.

    o   glm() now returns correct linear predictor and fitted values
    for zero-weight cases.

    o   scan(strip.white=TRUE) and readline() could give incorrect
    answers when an input field was empty or blank.

    o   dchisq() now behaves better for large values of its ncp parameter.

    o   besselJ(*, nu) is now ok for nu < 1; fix improves accuracy of
    other values by a few bits as well.

    o   The convergence criterion in the Fortran SVD code has been
    changed to circumvent spurious convergence errors when
    comparing extended-precision quantities on ix86 Linux (and
    some compilers and options).

    o   Rdindex now handles multi-line \title{}'s correctly.

    o   Add weights.glm() so that plot.lm() gets the correct weights from
    glm objects (the ones that go with deviance residuals not the
    working residuals).

    o   Printing a language object (e.g. a formula) in a list was
    clobbering the buffer holding the tag labels which then were
    lost in subsequent lines of printout (PR#746).

    o   aic in family() objects is now computed more precisely using
    dxxxx(log=TRUE).  For binomial fits the value was wrong for
    models specified via proportions and weights (PR#796).

    Avoid NaN values in aic where n = 0 occurs.

    o   Using non-integer numbers of successes in a binomial glm (as
    formerly in demo(lm.glm)) now gives a warning, as the theory and
    in particular the AIC value is not applicable.

    o   demo(is.things) works again {is.ALL() doesn't call methods
    anymore}.

    o   persp(*, nticks = nn) now works (better) for too small nn.
    persp(*, main = t1, sub= t2) now work (as always documented).

    o   Printing of attributes of a list element no longer precedes
    `attr' by the tag (PR#715).

    o   It is no longer possible to use df[foo] to select non-existent
    columns in df and so create an invalid object of class
    "data.frame" (PR#698).

    o   edit.data.frame() was trying to set row.names on a list, which
    is no longer allowed.

    o   In the \link[pkg]{topic}, `pkg' was not being used if `topic'
    was found in a package in .lib.loc.  Now `pkg' is always used.

    o   plot() with small relative ranges gave an error, which e.g. broke
    library(ts); plot(stl(ts(rep(1:7, 15), freq = 7), s.win = "per"))

    o   Using scan() with a 0-length `what' argument now gives an error
    message rather than a segfault.

    o   Loading .Random.seed could extremely rarely and with some
    generators reject a value as NA_INTEGER when it was valid.

    o   save(..., ascii=TRUE) created broken files when saving strings
    containing special characters (ASCII code > 127).

    o   mean(c(1,NA,NA,NA)[-1], trim = .1, na.rm = TRUE) doesn't give an
    error anymore.

    o   As the supplied afm files are not properly in the ISOLatin1
    encoding, the font metric for "-" was wrong since 0.63 on a
    postscript/xfig device, and randomly wrong in 1.2.0.  It will
    now be correct (but the files remain incorrect: C45 is "minus"
    not "hyphen" and there are missing slots).

    Metrics for the duplicated characters (acute, dieresis,
    macron, space) in ISOLatin1 encoding are now correct for both
    occurrences: previously only one was present.

    o   The data-time functions try harder to guess the DST setting
    when the OS does not know, as some OSes were printing times
    in DST if this was marked as unknown.

    o   Setting par(pty=) after par(pin=) or par(plt=) was having no
    effect until the margins were set.

    o   Nested \describe sections in .Rd files have (again) nested
    indentation when converted to text.

    o   Concatenation of C strings by ## is no longer used, to avoid
    spurious warnings from some recent versions of gcc.

    o   dev.copy2eps() allows a `paper' argument to be passed to
    postscript().

    o   file.show() handles the case of zero files better.

    o   formula.lm uses object$formula if it exists, in preference
    to object$terms.

    o   strsplit() is corrected for an empty split pattern.

    o   kronecker() now does a better job when creating dimnames.

    o   Better handling of input dimnames in fourfoldplot().



        CHANGES IN R VERSION 1.2.0


NEW FEATURES

    o   There is a new memory management system using a generational
    garbage collector.  This improves performance, sometimes
    marginally but sometimes by double or more.  The workspace is
    no longer statically sized and both the vector heap and the
    number of nodes can grow as needed.  (They can shrink again,
    but never below the initially allocated sizes.)  See ?Memory
    for a longer description, including the new command-line
    options to manage the settings.

    o   values of `--min-nsize' up to 50M (2Gb on 64-bit Solaris) are
    allowed.

    o   A (preliminary) version of S4-like connections has been added,
    and most functions which take a "file" argument can now work
    with a connection as well as a file name.  For more details,
    see the chapter on Connections in the R Data Import/Export
    manual.

    o   New command-line option `--no-restore-history' implied by
    `--vanilla'.

    o   Command-line option `--no-restore' is now `--no-restore-data'
    and `--no-restore' implies `--no-restore-*' (currently `data'
    and `history').

    o   The more recent GNU regex from grep-2.4.2 is used.  This uses
    locale-based ordering for ranges on platforms with strcoll.

    o   The print routines now escape " (as \") in a character string
    only when it is printed as a quoted string.  This makes
    print(, quote=FALSE) and cat() consistent.


    o   The standard methods for add1() and drop1() now attempt to cope
    with missing values by using a subset of the data that is
    `cleaned' by na.action for the maximal model under consideration.

    o   anova() for 3 or more lm objects now behaves compatibly with S
    and anova.glmlist().  The old behaviour is still available by
    calling anovalist.lm() directly.

    o   anova() for multiple lm and glm objects no longer truncates the
    formula printed. There is much more extensive documentation
    for anova() methods.

    o   New method as.data.frame.table() for converting the array-based
    representation of a contingency table to a data frame containing
    the classifying factors and the corresponding counts.

    o   New function assocplot() for producing Cohen-Friendly
    association plots.

    o   autoload() accepts lib.loc and other arguments to library()

    o   bxp() has new argument `frame.plot', as plot.default().

    o   contour() now has `axes' and `frame.plot' args.

    o   contrasts(, FALSE) now always returns an identity matrix,
    to make model.matrix compatible with S. This affects models
    such as lm(y ~ o - 1) where o is an ordered factor.

    o   `where' command added to debug().

    o   demo(dynload) (which used the superseded call_R interface)
    has been removed.

    o   Class "dendrogram" in package mva providing general support
    for tree-like structures (plotting, cutting, ...).

    o   dev.copy2eps() and dev2bitmap() preserve the aspect ratio of the
    copied device if just one of `width' and `height' is specified.

    o   dump() has new argument append, argument `fileout' has been
    renamed to `file' (for consistency with all other functions).

    o   edit.default() now checks for an unset `editor' argument, and
    terminates with an error if the editor cannot be run.

    o   The `mode' argument of exists() and get() is interpreted
    as mode(x) rather than typeof(x), following S.

    o   New functions file.access() and file.info() for information on
    files on the user's file systems.

    o   New convenience function file.copy().

    o   file.show() allows `pager' argument to be an R function, and
    consequently, the `pager' option can be an R function.

    o   Formatting (and printing) of data.frames with complex objects is
    improved.  toString was added as a new function.

    o   format() has a new argument `justify' controlling the
    justification of character strings (and factors).

    o   Formula objects now have an environment and code manipulating
    them needs to take care to preserve it or set an appropriate
    environment.

    o   New function fourfoldplot() for producing fourfold displays of
    2 by 2 by k contingency tables.

    o   gc() now reports the space allocated, not the space free,
    since the total space is now variable.

    o   New primitive gc.time() to report on time spent in garbage
    collection.

    o   hclust() takes new argument `members' allowing dissimilarity
    matrices both for singletons (as until now) and clusters.

    o   help() has an additional `pager' argument which may be passed to
    file.show()  {useful for ESS fans}.

    o   There is now an R `Hershey' list object for Hershey vector font
    computations and documentation.

    o   hist() now returns a "histogram" object and calls the new
    function plot.histogram() for plotting.
    It now also allows character labels.

    o   if(*) now gives a more intelligible error message
    when "*" cannot be coerced to logical.

    o   inherits() is now an internal function and compatible with S.

    o   New function lag.plot() in package ts.

    o   legend() has a new argument pt.bg.

    o   The commands history can be loaded with loadhistory(), saved
    with savehistory() and displayed with history(), under Windows
    and under Unix using the readline or GNOME interfaces.

    o   mad() has new (logical) arguments "low" and "high" (the first
    giving S compatibility).

    o   New function manova() and summary method.

    o   Function mantelhaen.test() in package ctest now can deal with
    general I x J x K tables.  In addition, in the 2 x 2 x K case,
    it can also perform an exact conditional test of independence,
    and gives confidence intervals for the common odds ratio.

    o   model.frame() now uses the environment of its formula argument,
    rather than the parent environment, to evaluate variables not
    found in the data argument. See help(formula).

    o   mosaicplot() can now also create extended mosaic plots, which
    visualize the residuals from a log-linear model using color and
    outline.

    o   New utility function n2mfrow().

    o   nlm(check.analyticals = TRUE) now warns if the supplied
    gradient and/or hessian are of the wrong length.

    o   New function object.size() to give approximate memory allocation.

    o   optim() now checks the length of an analytical gradient at
    each evaluation.

    o   The L-BFGS-B method of optim() now support tracing, at several
    levels of detail.

     o  options(check.bounds = TRUE)  makes each vector extension
    by sub-assignment ("x[.] <- .") produce a warning.

   o    options(width) now admits to a limit (previously 200, now
    10000) and gives a more informative message if out of range
    (as it does now for digits and expressions).

    o   Function path.expand() to do tilde-expansion on file paths.
    This provides an interface to R_ExpandFileName, which is now
    a documented entry point.

    o   .Platform has new component "endian", useful for binary file
    manipulations.

    o   plot.function() and curve() now take xlim as default for (from,to)
    if the former is specified.

    o   plot.hclust() allows arguments main, sub, etc, and has non-empty
    defaults for these.

    o   plot.ts(x,y) now allows to suppress labels and lines;
    it is better documented.

    o   The postscript() driver now allows a user-specified family so,
    for example, one can use the same fonts in diagrams as in
    running text.

    o   The postscript() driver allows its prolog to be changed (by an
    expert) via object .ps.prolog.

    o   prop.table() and margin.table() now work with an empty `margin'.

    o   Formerly deprecated function provide() is now defunct.

    o   New functions read.delim() and read.delim2() to make it
    easier to read delimited files as Windows programs tend to
    create (usually TAB separated).

    o   New readLines() function to read a file line-by-line.

    o   New functions reshapeLong() and reshapeWide() emulating Stata's
    reshape command. These are still labeled experimental and
    might be improved (or removed) in later versions.

    o   row.names() and row.names<-() are now generic functions which
    call rownames() as their default method and have methods for
    class "data.frame".

    o   New function Rprof() for profiling R expressions under Unix.
    Configure with `--enable-R-profiling' (on by default) to make
    this operational.

    o   save(, oldstyle=TRUE) has been withdrawn.

    o   scan() and read.table() have a new argument `fill' which can
    be set TRUE to allow reading files with unequal number of
    fields per line. (Programs like Excel have a habit of creating
    such files when exporting.)

    o   scan() and read.table() have a new argument `blank.lines.skip'
    to allow blank lines to be read.

    o   scan() now reads empty character fields as "" not "NA" unless
    "" is included in na.strings.

    o   smooth() in package eda has a better default (3RS3R instead of
    3RSR) and more arguments, e.g. `twiceit' for some S
    compatibility and `kind = "3R"' for running medians of 3.

    o   strsplit() has a new argument `extended' controlling whether to
    use extended (the default) or basic regular expressions for
    splitting.

    o   Sys.getenv() becomes the preferred name for getenv(), which is
    now deprecated.

    o   New functions Sys.getlocale() and Sys.setlocale() to query and
    set aspects of the locale of the R process, and
    Sys.localeconv() to find the default decimal point, etc.

    o   New function Sys.info() for platform, host and user information.

    o   New function Sys.putenv() to set environment variables.

    o   New function Sys.sleep() to suspend execution for a while.

    o   Date-time support functions with classes "POSIXct" and
    "POSIXlt" to represent dates/times (resolution 1 second) in
    the POSIX formats.  Functions include Sys.time(), as.POSIXct(),
    strftime(), strptime(),  and methods for format, plot, c, ....
    There are conversion functions for objects from packages
    `date' and `chron'; unlike those packages these support
    functions know about time zones (if the OS does).

    o   tcltk package now has tkpager() which is designed to be used
    by file.show() and shows help pages etc. in separate text
    widgets.

    o   tcltk is now more careful about removing the objects
    representing widgets in the R workspace when the windows are
    destroyed (e.g. using window manager controls)

    o   tcltk package has had several canvas functions implemented.

    o   tcltk now wraps callbacks to R in a try() construct - the
    nonlocal return from R's error handling could bring the Tk
    system into a strange state.

    o   New demos for tcltk: tkfaq, tkfilefind, tkcanvas.

    o   termplot() now has an `ask' argument.

    o   terms() creates objects which now inherit from class "formula",
    so for example as.formula(terms.object) needs to be replaced by
    formula(terms.object).

    o   traceback() is now printed un-quoted and labelled by the
    frame number.

    o   New argument `recursive' to unlink(). The default behaviour on
    Unix is now that of rm -f, not rm -rf.  unlink() is now
    compatible across platforms.

    o   New functions write.ftable() and read.ftable() for writing out
    and reading in flat contingency tables.

    o   write.table() now quotes factor columns if quote=TRUE, and has
    a new argument `qmethod' to control the escaping of
    embedded quotes in character or factor columns.

    o   New function xtabs() providing a formula interface to cross
    tabulation.


    o   The R Data Import/Export Manual (`R-data.texi') has been added.

    o   The set of valid R names is now described (at last) in R-intro.

    o   The R Language Manual (`R-lang.texi') is now included and built
    in the same way as the other manuals.

    o   The R manuals (R-intro, R-exts, ...) are converted to HTML
    format (if the necessary texinfo tools are available) and
    linked into the top HTML help page.

    o   The header file `R.h' and those included from it are now usable
    with C++ code.

    o   New header file `R_ext/Boolean.h': Rboolean type with TRUE and
    FALSE enum constants.

    o   New header file `Rgraphics.h' to allow addons to use graphics
    structures.

    o   Recommended include file `Rmath.h' replaces `R_ext/Mathlib.h'.

    o   Bessel, beta and gamma functions are now documented as part of
    the API.  Undocumented entry points are no longer in the
    header files, and some are no longer visible.

    o   Calloc & Realloc failures now give size information.

    o   DESCRIPTION file in installed packages has a new `Built:' field
    giving build information (R version, platform, date).

    o   Much improved support for C++ code in add-on packages under Unix.
    New configure/build variables SHLIB_CXXLD and SHLIB_CXXLDFLAGS
    for specifying the command and flags needed for building shared
    libraries containing objects from a C++ compiler.
    Configure tries to get these right in typical cases (GNU tools
    and/or common platforms).
    C++ source suffixes `.cpp' and `.C' are now recognized in
    addition to `.cc'.

    o   Configure/build variables MAINLD and MAINLDFLAGS are renamed to
    MAIN_LD and MAIN_LDFLAGS for consistency with other MAIN_* vars,
    similarly for SHLIBLD and SHLIBLDFLAGS.

    o   Configure/build variable FLIBS now only contains the Fortran 77
    intrinsic and run-time libraries needed for linking a Fortran 77
    program or shared library (as determined by configure).  BLAS
    library detection was extended, with results saved to the Make
    variable BLAS_LIBS which is also available to add-on packages.

    o   R CMD build and check have been completely re-written in Perl.
    In addition to running examples, check now also checks the
    directory structure and control files, makes a temporary
    installation and runs LaTeX on the help pages.  Build has been
    reduced to cleaning, rewriting indices and creating tar files.

    The same files of Perl code are now also used under Windows.

    o   Add-ons for utilities like Perl or LaTeX have now a central
    place in $R_HOME/share. Migration of existing files might take
    a while, though.

    o   Preliminary support for building R as a shared library (`libR')
    under Unix.  Use configure with option `--enable-R-shlib' or do
    `make libR' in directory `src/main' to create the shared
    library.

    There is also a linker front-end `R CMD LINK' which is useful
    for creating executable programs linked against the R shared
    library.


BUG FIXES

    o   New "[.AsIs" function fixing PR#665.

    o   Effectively zero-length arrows() are omitted (rather than getting
    a system- and device-dependent arbitrary angle).

    o   barplot() passes "..." down to plot.window as documented.

    o   bxp(*) now obeys axes, xaxt, yaxt & ylim arguments (again?).

    o   contour(.., labcex = 0.6, .. , vfont = c("sans serif", "plain"), ..)
    now behaves as it has been documented, PR#740.

    o   D(.) is now more correctly documented, has second argument
    "name" instead of "namevec" (S compatibility), and gives a warning
    when name has not length 1.

    o   dbinom() and dpois() now use a the accurate algorithm of C.Loader;
    e.g., 1 - 2*sum(dbinom(0:999, 1999, 1/2)) is now zero instead of 1e-12.

    o   dbeta(), df(), dt(), dnbinom(), dgeom(), dgamma(),
    now all use the more accurate algorithms of dbinom() and dpois();
    More limits are computed correctly, e.g. dgeom(Inf,*), or
    dnbinom(*,*,p=0), and integer argument coercion is more consistent.

    o   dbeta(x, a,b) was giving NaN and Inf for large a and b.

    o   The math functions now (again!) return NaN (not NA)
    when the warning message says so (e.g. gamma(-1)).

    o   Paper sizes for "a4" and "legal" were interchanged in dev.print().

    o   Improved logic in the re-scaling of plots in dev.print().

    o   dev.print() and dev.copy2eps() now evaluate arguments as
    needed, so e.g. dev.print(file="junk.ps", horizontal=F) will
    work (previously =FALSE was needed).

    o   Now dweibull(0,1) == dexp(0) == 1 for consistency.

    o   Changing column modes from factor/character to numeric in
    edit.data.frame() (and fix() on data frames) is now allowed.

    o   gctorture() did return a value although documented not to, that of
    gcinfo.  It now returns the previous setting of gctorture.

    o   glm() now follows lm() in dropping unused factor levels when fitting.

    o   hist() now labels y-axis Density rather than Relative Frequency
    if freq=FALSE. The `intensities' component of the result is
    now deprecated and `density' should be used instead.

    o   hsv() truncation bug fixed (hopefully correctly this time...)

    o   library.dynam() no longer fails if the library appears more than
    once in lib.loc.

    o   lm.fit & lm.wfit (and hence lm) now give understandable error
    messages when there are 0 non-NA cases.

    o   match() does not always use "as.character()" anymore, fixes PR#741.

    o   Trimmed mean() and median() are no longer attempted for complex
    data.

    o   median() of empty vector is now NA.

    o   model.matrix() can now handle NAs in factors.

    o   model.matrix() now uses ":" not "." in labelling interaction
    terms, as terms() and the prototype do (PR#701).

    o   old.packages() now sorts version strings correctly by
    splitting at "." and "-".

    o   optim() has several new precautions against bugs, one of which was
    an infinite loop in the L-BFGS-B method if the derivatives
    were inaccurate (or if numerical derivatives were used and the
    problem was badly scaled).

    o   pairs() now obeys an "cex = " argument, and so does panel.smooth().

    o   pf() had a bug in the code handling large noncentrality
    parameters, going all the way back to the original Appl.Stat.
    algorithm (PR#752). This affected pbeta() as well.

    o   phantom(expr) in plotmath was used to calculate the bounding
    boxes, but the space was not actually used in typesetting the
    expression (PR#622).

    o   Problem with subsetting and "..." arguments in plot.formula()
    (PR#757)

    o   plot.lm() has a better default for its `ask' argument.

    o   plot.new() does *NOT* allow an argument (`ask') anymore.
    This has not been working (since ~ R 0.49), is low-level and
    hence should not break anything.

    o   plot.stl() now uses correct time labels for tickmarks of x-axes.

    o   pnorm() is somewhat more accurate (SIXTEN=16; PR#699) and
    `infinitesimally faster' in some cases, thanks to James M. Rath.

    o   Failure to open a postscript() device now gives diagnostic warnings.

    o   Zero the widths of undefined chars in postscript().

    o   Make font metrics local to the device in postscript(), so
    that simultaneous instances of postscript/xfig/bitmap can use
    different values for `family'.

    o   postscript() no longer uses the initclip operator, which
    was against the EPSF specifications and broke inclusion of
    graphics in some documents.

    o   In postscript(), cleaned up the internals to fix some
    cases where it was unclear whether a variable was meant to set
    or track device status.

    o   In some cases, emitted postscript violated the PS Document
    Structuring Conventions since some settings were carried over from
    one page to the next. (Caused breakage when turning
    antialiasing on/off in "gv" for instance.)

    o   The DocumentMedia comment in postscript() output was wrong,
    and the font DSC comments have been changed to the preferred
    style for DSC 3.0.

    o   pretty(x <- c(2.8,3))[1] - x[1] is now 0; in general,
    p <- pretty(x) now again fulfills  p[1] <= x[] <= p[length(p)].

    o   print.dist() {mva} now passes "..." to print.default.

    o   print.noquote("ABC") now works.

    o   prompt(fn) uses fn's "source" attribute if it has one.

    o   In prop.test(), the conf.int. now contains the estimate,
    even when that is 0 or 1 (PR#714).

    o   qnorm(0) and qnorm(1) now give -/+ Inf (again!).

    o   qt(*, log = TRUE) has a somewhat extended working range.

    o   Specifying col.names in read.table() now overrides any header
    line in the data file.

    o   Bug caused rpois() to generate biased variables in some cases.

    o   scale(scale=TRUE) gave all NAs in any columns containing NAs.

    o   eda's smooth() now computes Tukey's endrule properly.

    o   smooth.spline()'s GCV method now uses a more accurate formula
    when there are repeated x points.

    o   stars() fix (for 0/NA columns) PR#739.

    o   data(state)'s  state.abb  has been corrected to use official
    2-letter abbrevations.

    o   strsplit() had two memory leaks.

    o   tcltk interface now has better string quoting and no longer breaks on
    strings with unbalanced braces

    o   tcltk callbacks used to include a "%..." argument on the call if the
    callback function had a "..." argument. This is no longer the case.

    o   t.test() allows an input vector of length 2.

    o   Method dispatch in primitive functions such as `c' only
    occurred if they were called by that name, and in particular
    not if used in lapply or called as .Primitive("c").  It now
    always occurs.

    o   Method dispatching now forces evaluation of promises, so
    autoloading works with methods.

    o   c(0.099999994, 0.2) prints as 0.1 0.2 not 0.10 0.20 after
    correcting the precision calculations.

    o   Internal function fprec was overflowing for numbers around
    1e-305.

    o   [internal] .C("strsignif",*) doesn't use fixed size strings anymore
    (possible overflow).

    o   ff[i] <- r  now works for formula (and potentially other
    non-standard) objects ff.

    o   Errors in argument matching could cause crashes if error
    handler was used.

    o   NULL == ... now gives logical(0) instead of an error.
    This fixes a bug with e.g. apply(X,2,median, na.rm = TRUE) and
    all(NULL == NULL) is now TRUE.

    o   polylines with more than 64K pieces will now work on more X
    servers with the x11 device.

    o   R CMD Rd2dvi now again works with file names containing "_" or "$".

    o   R CMD Rdconv now handles .rd as input file extension correctly
    in all cases.

    o   make install could fail if there were no shared libraries to
    install (e.g. on vanilla Solaris).

    o   Rd's \describe{} constructs are converted better to text
    output, and \items of >=22 chars now work.

    o   Help aliases "(" and "{" now work, apart from as PDF hyperlinks.

    o   NA || x , NA && y (etc) do not give errors anymore, but work as
    the "|" and "&" operators and as documented (PR#749).

    o   Coercing lists to functions failed if argument list was empty.
    (PR#743). Also coercing functions to lists was wrong when the
    body was a single constant.


    **************************************************
    *                        *
    *          1.1 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 1.1.1


NEW FEATURES

    o   data(), example() and help() now search the loaded packages, then
    in their .lib.loc argument, the latter as pre-0.99.0.  See
    their help pages for the precise details.

    o   help() has a new argument `try.all.packages' set by the option
    "help.try.all.packages". If help is not found in the normal
    path and this is TRUE then a search is made of all packages
    for possible matches.

    o   Improved detection of Tcl/Tk at configure time.  One can either
    set configure variables TCLTK_CPPFLAGS and TCLTK_LIBS, or give
    the locations of the tclConfig.sh and tkConfig.sh scripts via
    the `--with-tcl-config' and `--with-tk-config' arguments.

    o   prop.trend.test() - test for trend in proportions.

    o   write.table() has new argument `dec' for setting the decimal
    separator (SPSS/Windows and probably others want commas in
    continental European locales).

    o   Advance warning: save(, oldstyle=TRUE) will no longer be
    available after this release.

    o   Symbols can now be coerced to expressions, making
    as.expression(quote(a)) work


BUG FIXES

    o   data() now works correctly with data-only packages.

    o   Standalone Mathlib can now be made if builddir = srcdir
    (working around an undocumented autoconf `feature').

    o   Check for pre-1.1.0 saved object .Postscript.options in postscript().

    o   More stringent checks on the arguments to dataentry().

    o   Specifying xlab as an expression now works again.

    o   dev2bitmap() works again.

    o   Numerous uses of `F' and `T' replaced by `FALSE' and `TRUE'.

    o   bxp() now extends `ylim' properly when notch=TRUE (and a
    notch is more extreme than other stats).
    Also, varwidth is okay again, see example(split).

    o   Making manuals without making R first now works.

    o   princomp() now works when some eigenvalues are tiny negative
    values (which are treated as zero).

    o   SearchEngine.html was not installed when builddir != srcdir.

    o   lm.{w}fit(x,y,..) now work when x has no column names.
    These functions no longer add and subtract n zeros when there's no
    offset.  They are now documented somewhat.

    o   Some HTML files used src="R.css" instead of href="R.css" causing
    some browsers to ignore the style sheet.  `R.css' now passes the
    w3c CSS validator and gives the intended results with Internet
    Explorer.  `R.css' is copied to all library directories.

    o   The data editor no longer crashes when entries contain more than
    45 characters, and works correctly if the field width is
    larger than the window width.

    o   Recording plots used to ignore graphics events before the last
    replay.

    o   plot(), points(), etc now give warnings and error for invalid
    `type' arguments.

    o   matplot() now allows the same `type's as plot().

    o   dev.print() and dev.copy2eps() now work when called from a
    function, such as dev2bitmap().

    o   anova.glmlist works correctly with decreasing models and
    test="F" now gets the correct residual df.

    o   Functions created (not just edited) with fix() now get
    environment .GlobalEnv not the body of fix.

    o   Rdconv was ignoring OS-specific parts of the files with some
    Perl versions. (This did not affect installing packages.)

    o   R CMD check  did not run more than one specific test.

    o   keyword2html omitted items when changing up a level, so
    the HTML search engine keyword list was missing `data' and
    `utilities', for example.

    o   aov() with multiple strata was setting some spurious options.

    o   Under UNIX, help(, htmlhelp=TRUE) uses the linked copy of the
    HTML file in ~/.R if help.start() has been run recently in the
    session, so the style sheet and hyperlinks should now be found.

    o   dotplot's arguments `color' and `lcolor' are now interpreted
    correctly if they are vectors of length > 1.  (It is not clear
    from the help that this was ever intended.)

    o   Error messages of e.g. do.call() are now spaced properly.

    o   summary(.matrix | .data.frame) now respects a `digits' argument.

    o   scan() recognises quoted strings by default again. The default
    for the quote= argument got set wrongly when it was
    introduced. They are, however, turned off when using sep="\n",
    since you probably want unprocessed lines of text in that
    case.

    o   fixed buglet where scan() would double a character following a
    terminating quote symbol.

    o   kmeans was failing on vector inputs (although not documented
    to accept such).

    o   fixes in predict.lm (John Maindonald)

    o   NCOL got confused by 1D arrays, which it now reports have 1
    column rather than NA.

    o   rep(1:2,c(1,-1)) segfaulted (PR 619)

    o   x[["a"]] <- y rejected some object types for y (e.g.
    expressions) which were accepted in x$a <- y (PR

    o   data editor froze on empty data frame, and more generally on
    zero-length cols.

    o   data editor did not handle factors properly in numeric-edit
    mode.

    o   table() misbehaved on empty (or all-NA) data

    o   data editor now handles empty data frames, allowing data entry
    to begin with  dd<-edit(data.frame())  or
    dd<-data.frame() ; fix(dd)

    o   plotting math expressions when specified as a LANGSXP or
    SYMSXP didn't work in several cases: plot(1,main=quote(a))
    would segfault and there were problems with title(), text(),
    and mtext() as well.

    o   optim(): no more segfault for REPORT=0; corrected error msg; ..
    maxit default is now okay also when "L-*-B" is chosen automatically.
    Using (abs|rel)tol with that method now gives a warning.

    o   text() did not respect its `xpd' argument, the default of which
    is still to use par("xpd"); gave a wrong error message in one case.

    o   polygon() had an incorrect documentation and silently disregarded
    "..." arguments.

    o   predict.loess was giving incorrect results (those for the
    non-robust fit) in the case of a fit with family = "symmetric"
    and surface = "direct" and prediction with se = FALSE.  The
    fitted values were missing for such fits.

    o   Better rendering of ^ and ~ in latex in \code{} (and they should
    not appear outside code and verbatim sections).

    o   Fixed unterminated string problem with lty (PR 584)

    o   Fixed scoping problem with quasi() family (PR 614)


        CHANGES IN R VERSION 1.1.0


NEW FEATURES


    o   There are several changes to the compile and link flags, and
    an attempt is made to use /usr/local/include and
    /usr/local/lib on all platforms.  Please read config.site and
    INSTALL if you have set flags for previous versions.

    o   New package tcltk, providing interface and language bindings for
    the Tcl/Tk graphical interface toolkit. This makes it possible
    to do some rather nifty things with buttons and entry fields
    and so forth in very few lines of code.  It is still somewhat
    experimental so don't go churning out thousands of lines of
    GUI code and then blame us for changing the API later on. It does
    not currently work with GNOME because of event loop differences.
    See demo(tkttest) and demo(tkdensity).

    o   Internally using global R_Expressions [getOption("expressions")]
    dramatically speeds up eval.c.

    o   The X11 support code is now in a separate shared library, and
    option `--gui=none' will run R without X11 support (and so x11()
    and data.entry() will not be available).  R BATCH sets --gui=none.

    o   Configuring without X (e.g., `--without-x') does work (at last).
    configure will report if X11 support has been included.
    If R has been built without X11 the default GUI is "none".

    o   GNOME support files have moved from src/gnome to src/unix/gnome
    and the code is now in a separate shared library.  Both
    `--gui=GNOME' and `--gui=gnome' are allowed.

    The main graphics device for use with GNOME is now called
    gtk(); the x11() device from the X11 support can also be used.
    (The gnome() device remains unreliable, and is by default
    unavailable.)

    All GNOME support now installs in R_HOME/gnome to make it
    easier to package separately.

    o   Command line option `-g' can now be used as short version of
    `--gui'.

    o   The "keep.source" option now defaults to interactive()
    instead of TRUE.

    o   Many interface enhancements to the data editor, which now uses
    (by default) variable-width columns and much less redrawing.
    See ?dataentry for a description of the current user interface.

    o   R under Unix can be sent SIGUSR1 to quit and save or SIGUSR2
    to quit and save without running .Last and on.exit
    expressions.

    o   Added ability to use the graphics device driver to perform
    horizontal justification of strings.  This is used in the
    x11(), windows(), postscript() and xfig() drivers and will
    lead to more accurate placing of centred or right-justified
    strings.

    o   Formulas created with the tilde operator now contain the
    environment in which they were created.  The environment of a
    formula can be accessed with the environment() function.  This
    facility is still experimental; it is intended to help
    simplify writing modeling functions.


    o   $ and $<- are now generic functions.

    o   add1.glm() and drop1.glm() now allow test="F", which is
    appropriate for models with estimated dispersion.

    o   approx() and approxfun() have an extra argument ties= to control
    handling of tied x values

    o   New function as.symbol(), identical to as.name().
    is.symbol and is.name() are now identical by definition;
    before, they were equivalent by implementation.

    o   attach() now works for R data files.

    o   axis() now has additional arguments.  The argument "tick" turns
    tick drawing on and off (on by default).  Additional arguments
    line, pos, and outer; allow more flexible placement of axes.

    o   New device bitmap(), like dev2bitmap() but a stand-alone device.

    o   boxplot() returns (and bxp() accepts) a list containing the summary
    statistics for the boxes (changed from a list of elements one for
    each box).

    o   file="|cmd" is allowed (to pipe output to `cmd'), under cat()
    and postscript() and functions that use them (for example, write()).

    o   New functions chartr() for general-purpose character translation
    in character vectors, and tolower() and toupper() for converting
    alphabetic characters to lower-case and upper-case, respectively.

    o   coplot() has new option subscripts=TRUE to pass the indices
    to the panel function.

    o   cor.test(method="spearman") might give slightly better P values,
    since it now relies on pnorm() (with proper tail) instead of alnorm.

    o   dbinom() works for size=0 (where appropriate).

    o   New trivial function det() for determinant computation
    (the example in ?qr did not seem to be sufficient).

    o   If present in the package, a function .Last.lib will be called
    when a package is detached.

    o   New function dev.copy2eps() copies the current device to an eps
    file, taking the size from that of the device.

    o   dev.print() now prints if possible, and takes the size from
    the device being copied.

    o   edit() or fix() applied to a numeric or character matrix now
    uses the data editor, if available.

    o   edit.data.frame() edits factors in character mode by default,
    and will allow the row names to be edited, by default if they
    are not 1:nrow(name).

    o   expand.model.frame() for adding variables to a model frame.

    o   The handling of estimating the dispersion in glm() fits has been
    changed for consistency across methods.  For `binomial' and
    `poisson' families the dispersion is always taken as one.
    For `quasibinomial' and `quasipoisson' families it is always
    estimated (by residual chisq/residual df).

    o   Using "~" when specifying lib.loc (for example to help() or
    library()) is now supported.

    o   image() has a new argument "breaks" and divides the zlim range
    into equal-length intervals.

    o   install.packages() and update.packages() have a new `destdir'
    argument for the directory to which to download the packages.

    o   New function interaction.plot().

    o   Internal cleanup of lapply()'s C code.

    o   library(), require(), and sys.source() have a new argument
    ` keep.source = getOption("keep.source.pkgs") '.
    This defaults to false, but can be set (also for base) to true
    by setting the environment variable R_KEEP_PKG_SOURCE to "yes"

    o   Par `lty' in lines() now defaults to par("lty") not "solid",
    for consistency with plot().

    o   list.files() [aka dir()]: path argument has now default = "."

    o   locator() allows inline graphics parameters (for the plotting
    it does with type !="n").

    o   New function max.col() from MASS, thanks to V&R.

    o   nlm() can now take additional arguments in ... to be passed to f.

    o   par() returns elements in alphabetical order, and "ann" as a
    logical (not integer).

    o   .Platform has a new component GUI specifying the GUI in use
    (which might be "unknown").

    o   plot.formula() searches for a plot method for the left-side of the
    formula before searching for a plot method for the right-side of
    the formula.

    o   New bitmap devices png() and jpeg() on Unix/X11. (They were
    already available on Windows.)

    o   postscript(print.it=TRUE) now does print the plot file.
    postscript(onefile=FALSE) now produces multiple separate files.

    o   provide() is now deprecated.

    o   New function quade.test() in package ctest for the Quade test
    for unreplicated blocked data.

    o   quantile(x, probs) now propagates NA/NaN in probs.

    o   Option for user-supplied normal random generator in RNGkind().

    o   read.fwf() is no longer dependent on Perl.

    o   New arguments check.names and strip.white to read.table() and
    friends.

    o   Functions recordPlot() and replayPlot() save a plot to an R
    variable and replay it.

    o   residuals.lm() also allows types "partial" and "response"
    as residuals.glm().

    o   Added col option to rug().

    o   scale() is now generic with scale.default method.

    o   New "call. = TRUE" argument for stop(), and
    new function  stopifnot(.) -- to be used in "make check".

    o   str(.) has a new optional argument "nchar.max = 128" and truncates
    strings longer than that.  Also eliminated a few minor glitches.

    o   The symbols() function is now implemented.

    o   New function termplot() with similar functionality to
    plot.gam() in S.

    o   The parameters to title() which specify the labels may now
    be specified as a list containing the parameter together
    with additional modifying par values (cex, col, font) which
    affect just that label.

    title() also has new line and outer arguments which provide
    functionality which matches that in axis().

    o   New functions which.min() amd which.max().

    o   The Wilcoxon and Ansari-Bradley tests now optionally return
    confidence intervals for the location or scale parameter of
    interest.

    o   New argument `extend' to window() (and updated documentation).

    o   write.table() can now write CSV files in the format Excel expects.

    o   New device xfig() to produce XFig files.

    o   New xyz.coords() utility function to be used for 3D highlevel plots.


    o   R CMD build now also works for package bundles.

    o   Rdconv allows an output file to be specified.

    o   Initial support for C++ source code in add-on packages under Unix.

    o   Mathlib can be built as a standalone library: see directory
    src/nmath/standalone.



BUG FIXES

    o   The Mathlib functions now use exactly the same values for
    +/-Inf, NaN and NA as the rest of R in all cases.

    o   Improved automatic remaking.


    o   option keep.source was not set correctly before the base
    package was loaded: some functions in the base package had
    the source kept.  The memory usage should now be reduced by ca 1Mb.

    o   The base file is no longer closed twice in case of a syntax
    error (caused segfault on Linux).

    o   The X11 data editor scrolls cells and so allows more than ca 10
    characters to be entered.  Several interface/redraw errors
    have been corrected.

    o   Using `destroy' or `close' on the X11 data editor window killed R.

    o   Multiple `destroy' events or X11 errors killed R (PR#400).

    o   Warnings from .Last or closing devices are printed where
    sensible (e.g. not on GNOME nor Windows console).

    o   add1.glm() and drop1.glm() gave incorrect results in the "Chisq"
    test option when the scale was estimated.

    o   aggregate() applied to multiple time series lost the series names.

    o   as.hclust.twins() in package mva sets method=NA (PR#538)

    o   ave(1:3) now works (as when is.null(list()) was TRUE years ago).

    o   barplot(c(0.1, 0.5, 1), ylim = c(0, 1)) now works: [xy]axs="i"
    and exact limits now labels the limits.

    o   by() works when embedded in a function.

    o   The xlab and ylab in coplot() now set the x and y axis labels,
    as in S.  Supply a second component to label the conditioning
    variables.

    o   count.fields() and read.table() now have the same default
    quote = "\"'".

    o   guard against segfault in dataentry(x<-1, Modes = plot), and the like.

    o   dev.print() now checks if the current device is replayable
    (and not all devices were, despite the documentation).

    o   The gtk() device (formerly x11() under GNOME) was incorrectly
    returning character information, so in particular `*' was
    wrongly centred as a plotting symbol.

    o   is.qr() checks not only for $qr, but also for $rank & $qraux, PR#482.

    o   library() will not leave the package on the search path if
    loading fails (including if .First.lib fails).

    o   lqs(, method="S") rounded k0 on passing to C.

    o   na.omit() now handles vectors and matrices as well as data
    frames.

    o   plot.lm(*, which=4) now works thanks to Marcel Wolbers.

    o   postscript() now uses the options "papersize" and "printcmd"
    rather than the values of R_PAPERSIZE in the environment
    and R_PRINTCMD found at build time.

    o   postscript() now makes use of kerning information in computing
    string widths.

    o   M <- 2 ; print(ppr(.., nterms = M))  now works.

    o   ppr(, sm.method="spline") had an internal limit of 250 rows:
    this is now 2500 and is now checked.

    o   predict.lm(,type="terms",se.fit=T) gives correct standard errors
    (thanks to John Maindonald)

    o   print.ftable() now returns its argument.

    o   proj.aovlist() now gives a non-zero result for strata with
    no fitted effects.

    o   qnorm() is now based on AS 241 instead of AS 111, and should
    give precise results up to 16 digits precision.

    o   rgeom(, prob=1) gives 0 rather than NA

    o   rpois(*, lambda) now works when lambda > .Machine$integer.max

    o   strheight("X") with no active device would segfault.

    o   system() is now (partially) functional on systems without popen.

    o   Fixed bug in undoc() - requiring new `chdir' argument
    for sys.source()

    o   Fixed problem in usemethod() where CAR(call) could be non-symbol.

    o   The x11() device no longer crashes R if a font is not found,
    and will warn if there is a size mismatch of more than 10%.


    o   R CMD Rd2dvi now sorts files alphabetically.

    o   Rdconv now reports if an invalid type is specified (rather
    than using latex).

    o   Support for Rd conversion to Sd has been reinstated (it was
    unintentionally removed along with nroff-based conversion to text).

    o   \deqn{} fix in R-exts.texi, PR#523.


    o   Linpack.h now makes clear which routines are actually in R.

    o   Entry-point "pythag" is now loaded even though it is not
    necessarily used in R.  Where "hypot" is available pythag
    just calls hypot after checking for special values.

    o   call_R will now correctly handle character vector arguments.

    o   The alphabet links on top of HTML function indices with more than
    100 entries now contain only the letters which are really
    first letters of at least one entry.



        CHANGES IN R VERSION 1.0.1


BUG FIXES

    o   pgamma(Inf, 1, Inf) gives NaN instead of an infinite loop, PR#462.

    o   Inline xaxt and yaxt parameters were ignored.

    o   Setting the normal RNG was not always recorded in .Random.seed.

    o   \link[pkg]{foo} links were wrongly translated to HTML.

    o   Direct use of show.data() with no args was incorrect.

    o   print.coefmat does sensible things if options
    show.signif.stars and show.coef.Pvalues are not set correctly.

    o   x <- list(a=1, b=2) ; x["c"] <- 3 now (again) stretches the list
    rather than give a "subscript out of bounds" error.

    o   Models with response appearing on rhs could cause segfault.

    o   as.data.frame.matrix will now ensure the columns added have
    non-empty names.

    o   Some (strange?) model formulas could cause memory corruption
    (thanks to John Fox for digging this up).

    o   Matrix multiplication in cases such as 1:3 %*% t(1:2) is accepted.

    o   Options --nsize and --vsize to R CMD check gave warning
    messages and only one was accepted.

    o   The autoload()s during startup are "wrapped" such
    that conflicting objects don't break startup.

    o   cor.test(*, meth="spear") {pkg ctest} now works (again) for n <= 6

    o   scale = 0 in biplot.princomp() now works.

    o   The scores given by princomp(, cor = TRUE) and the rotated
    values given by prcomp(, scale = TRUE) are now computed from the
    scaled data.

    o   barplot now works okay when length(col) < length(legend.txt),
    thanks to Ben Bolker, PR#494.

    o   legend has new args `ncol' and `horiz' thanks to Ben Bolker.

    o   kernel("dirichlet", m=10, r=1) in package ts now works.

    o   save/restore of  environments containing "..." object should
    work now.

    o   qhyper should work better for large N.

    o   qbinom(1-1e-16, n, p) , qpois() and qnbinom(*) looped for ever.

    o   as.matrix.data.frame no longer escapes quotes in character
    vectors.

    o   R would try to save the readline history at the end of
    non-interactive sessions, causing crashes on some systems.

    o   Subtle roundoff problems caused arrays to be allocated short in
    loess.c in some cases, leading to segfaults.

    o   Document that postscript(file="") sends the output directly to
    the default printer.

    o   ts.union(, dframe=TRUE) had a long-standing typo, now works.

    o   save(, ascii=TRUE) had problem with octal escapes in strings
    when there was a digit following, should work now.

    o   matpoints() and matlines() now also have a "type =" argument
    with defaults "p" and "l", respectively; PR#506.

    o   The new Perl 5.6 broke our version-detection mechanism. A
    better one has been inserted.

    o   X11(colortype="gray") failed to work on 8-bit screens, and
    more generally when colour allocation failed X11() did not
    revert to monochrome correctly (and did not say it was doing so).

    It was not documented in the help that to change the X11
    colortype one needs to close all open X11 devices.

    o   loessf.f in modreg defined DSIGN unecessarily and this caused
    a conflict on AIX.

    o   plot.lm will now work with rlm objects.

    o   start and end behave better when the frequency is not an integer.

    o   aggregate.ts is closer to S-PLUS's aggregate.rts.

    o   quantile(x) now works when x contains several +/- Inf's.

    o   guard against segfault in dataentry(0,0) (or, more
    realistically, when mistyping data.entry(x,y) as
    dataentry(x,y))

    o   fixed special cases that caused bad behavior for methods,
    particularly with nested use of primitive operators when
    setMethod had been called.  Also fixed some situations that
    prevented caching methods (and therfore slowed down dispatch
    seriously) in similar cases.


    *********************************************************
    *                           *
    *   News of 1.0.0 and earlier is in file `ONEWS'    *
    *                           *
    *********************************************************