The R Project SVN R

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


        CHANGES IN R VERSION 2.7.0


NEW FEATURES

    o   identify() recycles its 'labels' argument if necessary.

    o   New function as.octmode() to create objects such as file
    permissions.

    o   dir.create() has a new argument 'mode', used on Unix-alikes
    (only) to set the permissions on the created directory.

    o   New function Sys.chmod(), a wrapper for 'chmod' on platforms
    which support it.

    o   as.POSIXlt() is now generic, and it and as.POSIXct() gain a
    '...' argument.  The character/factor methods now accept a
    'format' argument (analogous to that for as.Date).

    o   as.Date(), as.POSIXct() and as.POSIXlt() now convert numeric
    arguments (days or seconds since some epoch) provided the
    'origin' argument is specified.

    o   New higher-order function Negate().

    o   PCRE has been updated to 7.4.

    o   In addition to warning when 'pkgs' is not found,
        install.packages() now reports if it finds a valid package
        with only a case mismatch in the name.


DEPRECATED & DEFUNCT

    o   In package installation, SaveImage: yes is defunct and
    lazyloading is attempted instead.

    o   $ on an atomic vector or S4 object is now defunct.

    o   Partial matching in [[ is now only performed if explicitly
    requested (by exact=FALSE or exact=NA).


BUG FIXES

    o   Place holder



    **************************************************
    *                        *
    *          2.6 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 2.6.0 patched


BUG FIXES

    o   The "data.frame" and "factor" methods for [[ now support
    the 'exact' argument introduced in 2.6.0.

    o   iconv() is again able to translate character strings with
    embedded nuls (such as those in UCS-2).

    o   plot.lm() gains a new argument 'cex.caption' to allow the
    size of the captions to be controlled.

    o   new.packages() and update.packages() failed when called on an
    empty library, since old.packages() threw an error.
    old.packages() now returns NULL (as documented) in that case.



        CHANGES IN R VERSION 2.6.0


SIGNIFICANT USER-VISIBLE CHANGES

    o   integrate(), nlm(), nlminb(), optim(), optimize() and uniroot()
    now have '...' much earlier in their argument list.  This
    reduces the chances of unintentional partial matching but
    means that the later arguments must be named in full.

    o   The default type for nchar() is now "chars".  This is almost
    always what was intended, and differs from the previous default
    only for non-ASCII strings in a MBCS locale.  There is a new
    argument 'allowNA', and the default behaviour is now to throw
    an error on an invalid multibyte string if type="chars" or
    type="width".

    o   Connections will be closed if there is no R object referring to
    them.  A warning is issued if this is done, either at garbage
    collection or if all the connection slots are in use.


NEW FEATURES

    o   abs(), sign(), sqrt(), floor(), ceiling(), exp() and the gamma,
    trig and hyperbolic trig functions now only accept one
    argument even when dispatching to a Math group method (which
    may accept more than one argument for other group members).

    o   abbreviate() gains a 'method' argument with a new option
    "both.sides" which can make shorter abbreviations.

    o   aggregate.data.frame() no longer changes the group variables
    into factors, and leaves alone the levels of those which are
    factors.  (Inter alia grants the wish of PR#9666.)

    o   The default 'max.names' in all.names() and all.vars() is now
    -1 which means unlimited.  This fixes PR#9873.

    o   as.vector() and the default methods of as.character(),
    as.complex(), as.double(), as.expression(), as.integer(),
    as.logical() and as.raw() no longer duplicate in most cases
    where the object is unchanged.  (Beware: some code has been
    written that invalidly assumes that they do duplicate, often
    when using .C/.Fortran(DUP=FALSE).)

    o   as.complex(), as.double(), as.integer(), as.logical() and
    as.raw() are now primitive and internally generic for
    efficiency. They no longer dispatch on S3 methods for
    as.vector() (which was never documented).  as.real() and
    as.numeric() remain as alternative names for as.double().

    expm1(), log(), log1p(), log2(), log10(), gamma(), lgamma(),
    digamma() and trigamma() are now primitive.  (Note that logb()
    is not.)

    The Math2 and Summary groups (round, signif, all, any, max, min,
    summ, prod, range) are now primitive.

    See under METHODS PACKAGE below for some consquences for S4 methods.

    o   apropos() now sorts by name and not by position on the search
    path.

    o   attr() gains an 'exact = TRUE' argument to disable partial
    matching.

    o   bxp() now allows 'xlim' to be specified. (PR#9754)

    o   C(f, SAS) now works in the same way as C(f, treatment), etc.

    o   chol() is now generic.

    o   dev2bitmap() has a new option to go via PDF and so allow
    semi-transparent colours to be used.

    o   dev.interactive() regards devices with the displaylist enabled
    as interactive, and packages can register the names of their
    devices as interactive via deviceIsInteractive().

    o   download.packages() and available.packages() (and functions
    which use them) now support in 'repos' or 'contriburl' either
    file: plus a general path (including drives on a UNC path on
    Windows) or a file:/// URL in the same way as url().

    o   dQuote() and sQuote() are more flexible, with rendering
    controlled by the new option 'useFancyQuotes'.  This includes
    the ability to have TeX-style rendering and directional quotes
    (the so-called 'smart quotes') on Windows.  The default is to
    use directional quotes in UTF-8 locales (as before) and in the
    Rgui console on Windows (new).

    o   duplicated() and unique() and their methods in base gain an
    additional argument 'fromLast'.
    o   fifo() no longer has a default 'description' argument.

    fifo("") is now implemented, and works in the same way as file("").

    o   file.edit() and file.show() now tilde-expand file paths on all
    interfaces (they used to on some and not others).

    o   The find() argument is now named 'numeric' and not 'numeric.':
    the latter was needed to avoid warnings about name clashes
    many years ago, but partial matching was used.

    o   stats:::.getXlevels() confines attention to factors since some
    users expected R to treat unclass(<a factor>) as a numeric
    vector.

    o   grep(), strsplit() and friends now warn if incompatible sets
    of options are used, instead of silently using the documented
    priority.

    o   gsub()/sub() with perl = TRUE now preserves attributes from the
        argument x on the result.

    o   is.finite() and is.infinite() are now S3 and S4 generic.

    o   jpeg(), png(), bmp() (Windows), dev2bitmap() and bitmap() have
    a new argument 'units' to specify the units of 'width' and
    'height'.

    o   levels() is now generic (levels<- has been for a long time).

    o   Loading serialized raw objects with load() is now considerably
        faster.

    o   New primitive nzchar() as a faster alternative to
    nchar(x) > 0 (and avoids having to convert to wide chars in
    a MBCS locale and hence consider validity).

    o   The way old.packages() and hence update.packages() handle
    packages with different versions in multiple package
    repositories has been changed.  The first package encountered
    was selected, now the one with highest version number.

    o   optim(method = "L-BFGS-B") now accepts zero-length parameters,
    like the other methods.  Also, method = "SANN" no longer
    attempts to optimize in this case.

    o   New options 'showWarnCalls' and 'showErrorCalls' to give a
    concise traceback on warnings and errors.  showErrorCalls=TRUE
    is the default for non-interactive sessions.  Option
    'showNCalls' controls how abbreviated the call sequence is.

    o   New options 'warnPartialMatchDollar', 'warnPartialMatchArgs'
    and 'warnPartialMatchAttr' to help detect the unintended use
    of partial matching in $, argument matching and attr()
    respectively.

    o   A device named as a character string in options(device =) is now
    looked for in the grDevices name space if it is not visible
    from the global environment.

    o   pmatch(x, y, duplicates.ok = TRUE) now uses hashing and so is
    much faster for large x and y when most matches are exact.

    o   qr() is now generic.

    o   It is now a warning to have an non-integer object for
    .Random.seed: this indicates a user had been playing with it,
    and it has always been documented that users should only save
    and restore it.

    o   New higher-order functions Reduce(), Filter() and Map().

    o   [g]regexpr() gain an 'ignore.case' argument for consistency with
    grep().  (This does change the positional matching
    of arguments, but no instances of positional matching beyond
    the second were found.)

    o   relist() utility, an S3 generic with several methods, providing
    an 'inverse' for unlist(); thanks to a code proposal from
    Andrew Clausen.

    o   require() now returns invisibly.

    o   The interface to reshape() has been revised, allowing some
        simplified forms that did not work before, and somewhat
        improved error handling. A new argument 'sep' has been
        introduced to replace simple usages of 'split' (the old
        features are retained).

    o   rmultinom() uses a high-precision accumulator where available,
    and so is more likely to give the same result on different
    platforms (although it is still possible to get different
    results, and the result may differ from previous versions of R).

    o   row() and col() now work on matrix-like objects such as data
    frames, not just matrices.

    o   Rprof() allows smaller values of 'interval' on machines that
    support it: for example modern Linux systems support
    interval = 0.001.

    o   sample() now requires its first argument 'x' to be numeric
    (in the sense of is.numeric()) as well as of length 1 and >= 1
    before it is regarded as shorthand for 1:x.

    o   sessionInfo() now provides details about package name spaces
        that are loaded but not attached.  The output of sessionInfo
        has been improved to make it easier to read when it is
        inadvertently wrapped after being pasted into an email
        message.

    o   setRepositories() has a new argument 'ind' to allow selections
    to be made programmatically.

    o   showMethods() has a "smart" default for 'inherited' such that
    showMethods(<genfun>, incl = TRUE)  becomes a useful short cut.

    o   sprintf() no longer has a output string length limit.

    o   storage.mode<- is now primitive, and hence makes fewer copies
    of an object (none if the mode is unchanged).  It is a little
    less general than mode<-, which remains available.  (See also
    the entry under DEFUNCT below.)

    o   sweep() gains an argument 'check.margin = TRUE' which warns
    about mismatched dimensions.

    o   The mathematical annotation facility (plotmath) now recognises
        a symbol() function which forces the font to be a symbol font.
        This allows access to all characters in the Adobe Symbol
        encoding within plotmath expressions.

    o   For OSes that cannot unset environment variables, Sys.unsetenv()
    sets the value to "", with a warning.

    o   New function Sys.which(), an interface to 'which' on
    Unix-alikes and an emulation on Windows.

    o   On Unix-alikes, system(, intern = TRUE) reports on very long
        lines that may be truncated, giving the line number of the
        content being read.

    o   termplot() has a default for 'ask' that uses dev.interactive().

    It allows 'ylim' to be set, or computed to cover all the
    plots to be made (the new default) or computed for each plot
    (the previous default).

    o   uniroot(f, *) is slightly faster for non-trivial f() because it
    computes f(lower) and f(upper) only once, and it has new optional
    arguments 'f.lower' and 'f.upper' by which the caller can pass these.

    o   unlink() is now internal, using common POSIX code on all platforms.

    o   unsplit() now works with lists of dataframes.

    o   The vcov() methods for classes "gls" and "nlme" have migrated to
    package 'nlme'.

    o   vignette() has a new argument 'all' to choose between showing
        vignettes in attached packages or in all installed packages.

    o   New function within(), which is like with(), except that it
        returns modified versions back of lists and data frames.

    o   X11(), postscript() (and hence bitmap()), xfig(), jpeg(), png()
    and the Windows devices win.print(), win.metafile() and bmp()
    now warn (once at first use) if semi-transparent colours are
    used (rather than silently treating them as fully
    transparent).

    o   New function xspline() to provide base graphics support of
    X-splines (cf grid.xspline).

    o   New function xyTable() does the 2D gridding "computations" used by
    sunflowerplot().


    o   Rd conversion to HTML and CHM now makes use of classes, which
    are set in the stylesheets.  Editing R.css will change the
    styles used for \env, \option, \pkg etc.  (CHM styles are set
    at compilation time.)

    o   The documented arguments of '%*%' have been changed to be x
    and y, to match S and the implicit S4 generic.

    o   If members of the Ops group (the arithmetic, logical and
    comparison operators) and '%*%' are called as functions,
    e.g. '>'(x, y), positional matching is always used.  (It used
    to be the case that positional matching was used for the
    default methods, but names would be matched for S3 and S4
    methods and in the case of '!' the argument name differed
    between S3 and S4 methods.)

    o   Imports environments of name spaces are named (as
    "imports:foo"), and so are known e.g. to environmentName().

    o   Package 'stats4' uses lazy-loading not SaveImage (which is now
    deprecated).

    o   Installing help for a package now parses the .Rd file only
    once, rather than once for each type.

    o   PCRE has been updated to version 7.2.

    o   bzip2 has been updated to version 1.0.4.

    o   gettext has been updated to version 0.16.1.

    o   There is now a global CHARSXP cache, R_StringHash.  CHARSXPs
        are no longer duplicated and must not be modified in place.
        Developers should strive to only use mkChar (and mkString) for
        creating new CHARSXPs and avoid use of allocString.  A new
        macro, CallocCharBuf, can be used to obtain a temporary char
        buffer for manipulating character data.  This patch was
        written by Seth Falcon.

    o   The internal equivalents of as.complex, as.double, as.integer
    and as.logical used to handle length-1 arguments now accept
    character strings (rather than report that this is
    'unimplemented').

    o   Lazy-loading a package is now substantially more efficient (in
    memory saved and load time).

    o   Various performance improvements lead to a 45% reduction
    in the startup time without 'methods' (and one-sixth with -
    'methods' now takes 75% of the startup time of a default session).

    o   The [[ subsetting operator now has an argument 'exact' that
        allows programmers to disable partial matching (which will in
        due course become the default).  The default value is exact=NA
        which causes a warning to be issued when partial matching
        occurs.  When exact = TRUE, no partial matching will be
        performed.  When exact = FALSE, partial matching can occur and
        no warning will be issued.  This patch was written by Seth
        Falcon.

    o   Many of the C-level warning / error messages (e.g. from
    subscripting) have been re-worked to give more detailed
    information on either the location or the cause of the
    problem.

    o   The S3 and S4 Math groups have been harmonized.  Functions
    log1p(), expm1(), log10() and log2() are members of the S3
    group, and sign(), log1p(), expm1(), log2(), cummax(),
    cummin(), digamma(), trigamma() and trunk() are members of the
    S4 group.  gammaCody() is no longer in the S3 Math group.
    They are now all primitive.

    o   The initialization of the random-number stream makes use of the
    sub-second part of the current time where available.

    Initialization of the 1997 Knuth TAOCP generator is now done
    in R code, avoiding some C code whose licence status has been
    questioned.

    o   The reporting of syntax errors has been made more user-
    friendly.


METHODS PACKAGE

    o   Packages using 'methods' have to have been installed in R 2.4.0
    or later (when various internal representations were changed).

    o   Internally generic primitives no longer dispatch S4 methods on
    S3 objects.

    o   load() and restoring a workspace attempt to detect and warn on
    the loading of pre-2.4.0 S4 objects.

    o   Making functions primitive changes the semantics of S4
    dispatch: these no longer dispatch on classes based on types
    but do dispatch whenever the function in the base name space is
    called.

    This applies to as.complex(), as.integer(), as.logical(),
    as.numeric(), as.raw(), expm1(), log(), log1p(), log2(),
    log10(), gamma(), lgamma(), digamma() and trigamma(), as
    well as the Math2 and Summary groups.

    Because all members of the group generics are now primitive,
    they are all S4 generic and setting an S4 group generic does
    at last apply to all members and not just those already made
    S4 generic.

    as.double() and as.real() are identical to as.numeric(), and
    now remain so even if S4 methods are set on any of them.
    Since 'as.numeric' is the traditional name used in S4,
    currently methods must be exported from a NAMESPACE for
    'as.numeric' only.

    o   The S4 generic for '!' has been changed to have signature (x)
    (was (e1)) to match the documentation and the S3 generic.
    setMethod() will fix up methods defined for (e1), with a
    warning.

    o   The "structure" S4 class now has methods that implement the
        concept of structures as described in the Blue Book--that
        element-by-element functions and operators leave structure
        intact unless they change the length.  The informal behavior
        of R for vectors with attributes was inconsistent.

    o   The implicitGeneric() function and relatives have been added to
    specify how a function in a package should look when methods are
    defined for it.  This will be used to ensure that generic versions
    of functions in R core are consistent.  See ?implicitGeneric.

    o   Error messages generated by some of the functions in the methods
        package provide the name of the generic to provide more
        contextual information.

    o   It is now possible to use setGeneric(useAsDefault = FALSE) to
    define a new generic with the name of a primitive function
    (but having no connection with the primitive).


DEPRECATED & DEFUNCT

    o   $ on an atomic vector now gives a warning that it is 'invalid'.
    It remains deprecated, but may be removed in R >= 2.7.0.

    o   storage.mode(x) <- "real" and storage.mode(x) <- "single" are
    defunct: use instead storage.mode(x) <- "double" and
    mode(x) <- "single".

    o   In package installation, SaveImage: yes is deprecated in
    favour of LazyLoad: yes.

    o   seemsS4Object (methods package) is deprecated in favour of isS4().

    o   It is planned that [[exact=TRUE]] will become the default in
    R 2.7.0.


UTILITIES

    o   checkS3methods() (invoked by R CMD check) now checks the
    arguments of methods for primitive members of the S3 group
    generics.

    o   R CMD check now does a recursive copy on the 'tests' directory.

    o   R CMD check now warns on non-ASCII .Rd files without an
    \encoding field, rather than just on ones that are definitely
    not from an ISO-8859 encoding.  This agrees with the
    long-standing stipulation in 'Writing R Extensions', and
    catches some packages with UTF-8 man pages.

    o   R CMD check now warns on DESCRIPTION files with a non-portable
    Encoding field, or with non-ASCII data and no Encoding field.

    o   R CMD check now loads all the 'Suggests' and 'Enhances'
    dependencies to reduce warnings about non-visible objects, and
    also emulates standard functions (such as shell()) on
    alternative R platforms.

    o   R CMD check now (by default) attempts to latex the vignettes
    rather than just weave and tangle them: this will give a NOTE
    if there are latex errors.

    o   R CMD check computations no longer ignore Rd \usage entries for
        functions for extracting or replacing parts of an object, so S3
        methods should use the appropriate \method{} markup.

    o   R CMD check now checks for CR (as well as CRLF) line endings in
    C/C++/Fortran source files, and for non-LF line endings in
    Makefile[.in] and Makevars[.in] in the package 'src'
    directory.  R CMD build will correct non-LF line endings in
    source files and in the make files mentioned.

    o   Rdconv now warns about unmatched braces rather than silently
    omitting sections containing them.  (Suggestion by Bill
    Dunlap, PR#9649)

    Rdconv now renders (rather than ignores) \var{} inside \code{}
    markup in latex conversion.

    R CMD  Rdconv gains a --encoding argument to set the default
    encoding for conversions.

    o   The list of CRAN mirrors now has a new (manually maintained)
        column "OK" which flags mirrors that seem to be OK, only those
        are used by chooseCRANmirror(). The now exported function
        getCRANmirrors() can be used to get all known mirrors or only
        the ones that are OK.

    o   R CMD SHLIB gains arguments --clean and --preclean to clean up
    intermediate files after and before building.

    o   R CMD config now knows about FC and FCFLAGS (used for F9x
    compilation).

    o   R CMD Rdconv now does a better job of rendering quotes in
    titles in HTML, and \sQuote and \dQuote into text on Windows.


C-LEVEL FACILITIES

    o   New utility function alloc3DArray similar to allocMatrix.

    o   The entry point R_seemsS4Object in Rinternals.h has not been
    needed since R 2.4.0 and has been removed.  Use IS_S4_OBJECT
    instead.

    o   Applications embedding R can use R_getEmbeddingDllInfo() to
    obtain DllInfo for registering symbols present in the
    application itself.

    o   The instructions for making and using standalone libRmath have
    been moved to the R Installation and Administration manual.

    o   CHAR() now returns (const char *) since CHARSXPs should no
        longer be modified in place.  This change allows compilers to
        warn or error about improper modification.  Thanks to Herve
        Pages for the suggestion.

    o   acopy_string is a (provisional) new helper function that
        copies character data and returns a pointer to memory
        allocated using R_alloc.  This can be used to create a copy of
        a string stored in a CHARSXP before passing the data on to a
        function that modifies its arguments.

    o   asLogical, asInteger, asReal and asComplex now accept STRSXP
    and CHARSXP arguments, and asChar accepts CHARSXP.

    o   New R_GE_str2col() exported via R_ext/GraphicsEngine.h for
        external device developers.

    o   doKeybd and doMouseevent are now exported in GraphicsDevice.h.

    o   R_alloc now has first argument of type 'size_t' to support
    64-bit platforms (e.g. Win64) with a 32-bit 'long' type.

    o   The type of the last two arguments of getMatrixDimnames (non-API
    but mentioned in R-exts.texi and in Rinternals.h) has been
    changed to 'const char **' (from char **).

    o   R_FINITE now always resolves to the function call R_finite in
    packages (rather than sometimes substituting isfinite).  This
    avoids some issues where R headers are called from C++ code
    using features tested on the C compiler.

    o   The advice to include R headers from C++ inside extern "C" {}
    has been changed.  It is nowadays better *not* to wrap the
    headers, as they include other headers which on some OSes
    should not be wrapped.

    o   Rinternals.h no longer includes a substantial set of C headers.
    All but ctype.h and errno.h are included by R.h which is supposed
    to be used before Rinternals.h.

    o   Including C system headers can be avoided by defining NO_C_HEADERS
    before including R headers.  This is intended to be used from
    C++ code, and you will need to include C++ equivalents such as
    <cmath> before the R headers.


INSTALLATION

    o   The 'test-Lapack' test is now part of 'make check'.

    o   The 'stat' system call is now required, along with 'opendir'
    (which had long been used but not tested for).  ('make check'
    would have failed in earlier versions without these calls.)

    o   'evince' is now considered as a possible PDF viewer.

    o   'make install-strip' now also strips the DLLs in the standard
    packages.

    o   Perl 5.8.0 (released in July 2002) or later is now required.
    (R 2.4.0 and later have in fact required 5.6.1 or later.)

    o   The C function 'finite' is no longer used: we expect a C99
    compiler which will have 'isfinite'.  (If that is missing, we
    test separately for NaN, Inf and -Inf.)

    o   A script/executable 'texi2dvi' is now required on Unix-alikes:
    it is part of the texinfo distribution.

    o   Files texinfo.tex and txi-en.tex are no longer supplied in
    doc/manual (as the latest versions have an incompatible
    licence).  You will need to ensure that your texinfo and/or
    TeX installations supply them.

    o   wcstod is now required for MBCS support.

    o   There are some experimental provisions for building on Cygwin.


PACKAGE INSTALLATION

    o   The encoding declared in the DESCRIPTION file is now used as
    the default encoding for .Rd files.

    o   A standard for specifying package license information in the
        DESCRIPTION License field was introduced, see 'Writing R
        Extensions'.  In addition, files LICENSE or LICENCE in a package
        top-level source directory are now installed (so putting copies
        into the 'inst' subdirectory is no longer necessary).

    o   install.packages() on a Unix-alike now updates
    doc/html/packages.html only if packages are installed to
    .Library (by that exact name).

    o   R CMD INSTALL --clean now runs SHLIB --clean to do the clean
    up (unless there is a src/Makefile), and this will remove
    $(OBJECTS) (which might have been redefined in Makevars).

    R CMD INSTALL --preclean cleans up the sources after a
    previous installation (as if that had used --clean) before
    attempting to install.

    R CMD INSTALL will now run R CMD SHLIB in the 'src' directory
    if src/Makevars is present, even if there are no source files
    with known extensions.

    o   If there is a file src/Makefile, src/Makevars is now ignored
    (it could be included by src/Makefile if desired), and it is
    preceded by etc/Makeconf rather than share/make/shlib.mk.
    Thus the makefiles read are R_HOME/etc/Makeconf, src/Makefile
    in the package and then any personal Makevars files.

    o   R CMD SHLIB used to support the use of 'OBJS' in Makevars, but
    this was changed to 'OBJECTS' in 2001.  The undocumented
    alternative of 'OBJS' has finally been removed.

    o   R CMD check no longer issues a warning about no data sets
        being present if a lazyload db is found (as determined by the
        presence of Rdata.rdb, Rdata.rds, and Rdata.rdx in the 'data'
        subdirectory.


BUG FIXES

    o   charmatch() and pmatch() used to accept non-integer values for
    'nomatch' even though the return value was documented to be
    integer.  Now 'nomatch' is coerced to integer (rather than the
    result being coerced to the type of 'nomatch').

    o   match.call() no longer 'works' outside a function unless
    'definition' is supplied.  (Under some circumstances it used
    to 'work', matching itself.)

    o   The formula methods of boxplot, cdplot, pairs and spineplot
    now attach 'stats' so that model.frame() is visible where
    they evaluate it.

    o   Date-time objects are no longer regarded as numeric by
    is.numeric().

    o   methods("Math") did not work if 'methods' was not attached.

    o   readChar() read an extra empty item (or more than one) beyond
    the end of the source; in some conditions it would terminate
    early when reading an item of length 0.

    o   Added a promise evaluation stack so interrupted promise
        evaluations can be restarted.

    o   R.version[1:10] now nicely prints.

    o   In the methods package, prototypes are now inherited for the
    .Data "slot"; i.e., for classes that contain one of the basic
    data types.

    o   <data frame>[[i, j]] now works if 'i' is character.

    o   write.dcf() no longer writes NA fields (PR#9796), and works
    correctly on empty descriptions.

    o   pbeta(x, log.p = TRUE) now has improved accuracy in many cases,
    and so have functions depending on it such as pt(), pf() and
    pbinom().

    o   mle() had problems with the L-BFGS-B in the no-parameter case
    and consequentially also when profiling 1-parameter models
    (fix thanks to Ben Bolker).

    o   Two bugs fixed in methods that in involve the "..." argument in
    the generic function: previously failed to catch methods that
    just dropped the "..."; and use of callGeneric() with no arguments
    failed in some circumstances when "..." was a formal argument.

    o   sequence() now behaves more reasonably, although not
    back-compatibly for zero or negative input.

    o   nls() now allows more peculiar but reasonable ways of being called,
    e.g., with data=list(<uneven lengths>) or a model without variables.

    o   match.arg() was not behaving as documented when
    several.ok=TRUE (PR#9859), gave spurious warnings when 'arg'
    had the wrong length and was incorrectly documented (exact
    matches are returned even when there is more than one partial
    match).

    o   The data.frame method for split<-() was broken.

    o   The test for -D__NO_MATH_INLINES was badly broken and returned
    true on all non-glibc platforms and false on all glibc ones
    (whether they were broken or not).

    o   LF was missing after the last prompt when --quiet was used without
    --slave. Use --slave when no final LF is desired.

    o   Fixed bug in initialisation code in 'grid' package for
        determining the boundaries of shapes.  Problem reported
        by Hadley Wickham;  symptom was error message:
        "Polygon edge not found".

    o   str() is no longer slow for large POSIXct objects.  Its output
    is also slightly more compact for such objects; implementation
    via new optional argument 'give.head'.

    o   strsplit(*, fixed=TRUE), potentially iconv() and internal string
    formatting is now faster for large strings, thanks to report
    PR#9902 by John Brzustowski.

    o   de.restore() gave a spurious warning for matrices (Ben Bolker)
    
    o   plot(fn, xlim=c(a,b)) would not set "from" and "to" properly
    when plotting a function.  The argument lists to curve() and
    plot.function() have been modified slightly as part of the
    fix.

    o   julian() was documented to work with POSIXt origins, but did
    not work with POSIXlt ones.  (PR#9908)

    o   Dataset HairEyeColor has been corrected to agree with
    Friendly (2000): the change involves the breakdown of the
    Brown hair / Brown eye cell by Sex, and only totals over Sex
    are given in the original source.

    o   Trailing spaces are now consistently stripped from \alias{}
    entries in .Rd files, and this is now documented.  (PR#9915)

    o   .find.packages(), packageDescription() and sessionInfo()
    assumed that attached environments named "package:foo" were
    package environments, although misguided users could use such
    a name in attach().
    
    o   spline() and splinefun() with method = "periodic" could return
        incorrect results when length(x) was 2 or 3.
        
    o   getS3method() could fail if the method name contained a regexp
        metacharacter such as "+".
        
    o   help(<a character vector>) now uses the name and not the
    value of the vector unless it has length exactly one, so
    e.g. help(letters) now gives help on 'letters'.
    (Related to PR#9927)

    o   Ranges in chartr() now work better in CJK locales, thanks to
    Ei-ji Nakama.




    **************************************************
    *                        *
    *          2.5 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 2.5.1 patched

INSTALLATION

    o   doc/manual now includes the texinfo support file epsf.tex which
    basic TeX installations often omit.


BUG FIXES

    o   Attempting to do in-memory serialization of an object requiring
    more than 1Gb might have failed.

    o   Using formals<- on a function whose body was NULL worked
    incorrectly.  (PR#9758)

    o   logb() is now strictly a wrapper for log(), so if S4 methods
    are set on log(), logb() will also dispatch on them.

    o   conflicts(where=) did not work correctly. (PR#9760)

    o   log(x, base) was intended to handle complex 'base' even for
    real 'x', but there was a typo in the code to do so.

    o   Syntax errors would sometimes misreport the error context.

    o   qt(p, df=1) is now also correct for very small p. (PR#9804)
    qt(p, df=2) ditto; also is more accurate for  0 < |p - 0.5| << 1.
    qt(*, log.p=TRUE) now is finite and monotone (again!) where possible.

    o   Several functions including those making use of printCoefmat(),
    layout() and sortedXyData() now work correctly with
    non-default settings of options("OutDec").

    o   S4 method dispatch for group generics (and %*%) failed to pass
    arguments to methods as promises and so in some circumstances
    methods could change their arguments.  (Seen for the 'Math'
    group with package Matrix.)

    o   The print() method for "ts" now handles quarterly and monthly
    series which do not start at the beginning of a quarter or
    month respectively.

    o   Deserializing raw objects saved using save(..., ascii=TRUE)
        now works correctly.

    o   ISOLatin7 encoding for postscript/PDF has been corrected.
    (PR#9845)

    o   rbind(x,y) and cbind(x,y) did not dispatch properly when x and
        y had multiple S3-style classes.

    o   The workaround for seeking on > 2Gb files did not work correctly
    on Unix-alike 32-bit systems.  (PR#9883)

    o   We had identical(NaN, NA_real_) != identical(NA_real_, NaN),
    spotted by Petr Savicky.



        CHANGES IN R VERSION 2.5.1


NEW FEATURES

    o   density(1:20, bw = "SJ") now works as bw.SJ() now tries a larger
    search interval than the default (lower, upper) if it does not
    find a solution within the latter.

    o   The output of library() (no arguments) is now sorted by library
    trees in the order of .libPaths() and not alphabetically.

    o   R_LIBS_USER and R_LIBS_SITE feature possible expansion of
        specifiers for R version specific information as part of the
        startup process.

    o   C-level warning calls now print a more informative context,
    as C-level errors have for a while.

    o   There is a new option "rl_word_breaks" to control the way the
    input line is tokenized in the readline-based terminal
    interface for object- and file-name completion.
    This allows it to be tuned for people who use their space bar
    vs those who do not.  The default now allows filename-completion
    with +-* in the filenames.

    o   If the srcfile argument to parse() is not NULL, it will be added
        to the result as a "srcfile" attribute.

    o   It is no longer possible to interrupt lazy-loading (which was
    only at all likely when lazy-loading environments), which
    would leave the object being loaded in an unusable state.
    This is a temporary measure: error-recovery when evaluating
    promises will be tackled more comprehensively in 2.6.0.


INSTALLATION

    o   'make check' will work with --without-iconv, to accommodate
    building on AIX where the system iconv conflicts with
    libiconv and is not compatible with R's requirements.

    o   There is support for 'DESTDIR': see the R-admin manual.

    o   The texinfo manuals are now converted to HTML with a style
    sheet: in recent versions of makeinfo the markup such as @file
    was being lost in the HTML rendering.

    o   The use of inlining has been tweaked to avoid warnings from
    gcc >= 4.2.0 when compiling in C99 mode (which is the default
    from configure).


BUG FIXES

    o   as.dendrogram() failed on objects of class "dendrogram".

    o   plot(type ="s") (or "S") with many (hundreds of thousands)
    of points could overflow the stack.  (PR#9629)

    o   Coercing an S4 classed object to "matrix" (or other basic class)
    failed to unset the S4 bit.

    o   The 'useS4' argument of print.default() had been broken by an
    unrelated change prior to 2.4.1.  This allowed print() and
    show() to bounce badly constructed S4 objects between
    themselves indefinitely.

    o   Prediction of the seasonal component in HoltWinters() was one
    step out at one point in the calculations.

    decompose() incorrectly computed the 'random' component for a
    multiplicative fit.

    o   Wildcards work again in unlink() on Unix-alikes (they did not
    in 2.5.0).

    o   When qr() used pivoting, the coefficient names in qr.coef() were
    not pivoted to match.  (PR#9623)

    o   UseMethod() could crash R if the first argument was not a
    character string.

    o   R and Rscript on Unix-alikes were not accepting spaces in -e
    arguments (even if quoted).

    o   Hexadecimal integer constants (e.g. 0x10L) were not being parsed
    correctly on platforms where the C function atof did not
    accept hexadecimal prefixes (as required by C99, but not
    implemented in MinGW as used by R on Windows).  (PR#9648)

    o   libRlapack.dylib on Mac OS X had no version information and
    sometimes an invalid identification name.

    o   Rd conversion of \usage treated '\\' as a single backslash in
    all but latex: it now acts consistently with the other
    verbatim-like environments (it was never 'verbatim' despite
    the documentation).

    \code{\.} is now rendered as '\.' in all formats, as
    documented (it was not the case for latex conversion).

    codoc() (and checkDocStyle() and checkDocUsage()) now apply
    the same transformations to \usage as Rd conversion does,
    so {, % and \\ in strings in usages will now be related
    correctly to the help files.

    o   rbind() failed if the only data frame had 0 rows. (PR#9657)

    o   <a data.frame>[i, j] could sometimes select the wrong column
    when j is numeric if there are duplicate column names.

    o   sample(x, size, replace=TRUE, prob) had a memory leak if
    10000 < size <= 100000.

    o   x <- cbind(1:2); rownames(x) <- factor(c("A",NA))  now longer
    segfaults.

    o   R CMD BATCH no longer assumes Sys.unsetenv() is supported (it
    is not on older Solaris systems).

    o   median() returned a logical result when it was 'NA': it now
    returns an NA of appropriate type (e.g. integer or double).

    o   grep(fixed = TRUE, perl = TRUE) ignored 'fixed', although it
    was documented to ignore 'perl'
    Same for [g]regexpr and [g]sub.

    o   getNamespaceExports("base") works again.

    o   runmed(c(), 1) no longer segfaults.

    o   qr.coef(QR, b) failed for an LAPACK-produced QR if b was
    integer or for an over-determined system.

    qr.solve() for an under-determined system produces a
    solution with 0 and not NA for columns which are unused.

    o   segments() was not handling full transparency correctly in PDF.
    (PR#9694)

    Nor was arrows().

    o   callGeneric() inside a method with extra arguments {and hence
    currently defined via .local()} now works.

    o   [g]sub(fixed=TRUE, useBytes=FALSE) could substitute in the
    wrong place in an MBCS locale.

    gregexpr() could give incorrect answers in MBCS locales for
    perl = TRUE or fixed = TRUE (unless useBytes = TRUE).

    o   The legacy quartz() device no longer crashes in locator()
    if the user attempts to close the window.

    o   "CGGStackRestore: gstack underflow" warning is no longer shown
    in legacy quartz() device.

    o   formatC() now limits 'digits' to 50 to avoid problems in
    C-level sprintf in some OSes.

    o   seq.int(x, y, by=z) gave 'x' (and not an error) if
    0 > (y-x)/z > -1.

    o   promptClass() now lists methods, including those for generics in
    other attached packages.

    o   Connection-related functions such as readBin() no longer crash
    when supplied with a non-connection object.

    o   as.character.srcref() didn't handle bad srcref objects cleanly.

    o   predict.nls() no longer requires 'newdata' to contain exactly
    the variable names needed to fit the model: variables used on
    the LHS only are no longer required and further variables are
    allowed.

    o   plot.hclust() had a 'out by one' error, and ignored the last
    object when computing the window region (and could overrun
    arrays).

    o   deriv() was creating results with double (and not integer) dims.

    o   The unserialize code (e.g. as called by load()) looked for a
    function findPackageEnv() to set a saved package environment.
    This was missing, but is now supplied.

    o   [cr]bind could segfault when creating a list matrix result.
    (Reported by Martin Morgan.)

    o   besselI(x, nu, exp=TRUE) and besselY(x, nu) could give wrong
    answers for nu < 0. (Reported by Hiroyuki Kawakatsu.)

    o   [g]sub could confuse a trailing byte '\' for a backreference
    in MBCSs where '\' can occur as a trailing byte (not UTF-8 nor
    EUC-JP, but SJIS and the CJK character sets used on Windows).
    (PR#9751)



        CHANGES IN R VERSION 2.5.0


USER-VISIBLE CHANGES

    o   apropos(x) and find(x) now both only work for character 'x',
    and hence drop all non-standard evaluation behaviour.

    o   Data frames can have 'automatic' row names which are not
    converted to dimnames by as.matrix().  (Consequently, e.g.,
    t(.)  for such data frames has NULL column names.)  This
    change leads to memory reductions in several places, but can
    break code which assumes character dimnames for data frames
    derived from matrices.

    No existing R object is regarded as having 'automatic' row
    names, and it may be beneficial to recreate such objects via
    read.table() or data.frame().

    o   Using $ on an atomic vector now raises a warning, as does use
    on an S4 class for which a method has not been defined.

    o   The Unix-alike readline terminal interface now does
    command-completion for R objects, incorporating the
    functionality formerly in package 'rcompletion' by Deepayan
    Sarkar.  This can be disabled by setting the environment
    variable R_COMPLETION=FALSE when starting R (e.g. in
    ~/.Renviron).  (Note that when this is enabled, filename
    completion no longer works for file paths containing R
    operators such as '+' and '-'.)


NEW FEATURES

    o   abbreviate() no longer has an 8191 byte limit on the size of
    strings it can handle.

    o   abs(x) now returns integer for integer or logical arguments.

    o   apropos() has a new argument 'ignore.case' which defaults to
    TRUE, potentially matching more than previously, thanks to a
    suggestion by Seth Falcon.

    o   args(), str() and print() now give the argument lists of primitive
    functions.

    o   as.matrix() gains the '...' argument that several packages have
    assumed it always had (and S-PLUS has).

    o   Manipulation of integers as roman numerals via as.roman() in
    package utils.

    o   attr() no longer treats name = NA_character_ as meaning
    name = "NA".

    o   binom.test() now allows a 'fuzz' for calculated integer values
    in its x and n arguments.

    o   boxplot(*, notch = TRUE) now warns when notches are outside
    hinges; related to PR#7690.

    o   New function callCC() providing a downward-only version of
    Scheme's call with current continuation.

    o   capabilities() now has a "profmem" entry indicating whether
    R has been compiled with memory profiling.

    o   colnames<-() and rownames<-() now handle data frames explicitly,
    so calling colnames<- on a data frame no longer alters the
    representation of the row names.

    o   commandArgs() has a new 'trailingOnly' argument to be used in
    conjunction with --args.

    o   contour() now passes graphical parameters in '...' to axis()
    and box().

    o   New data set 'crimtab' on Student(1908)'s 3000 criminals.

    o   cut.default() has a new argument 'ordered_result'.

    o   .deparseOpts() has two new options: "keepNA" to ensure that
    different types (logical, integer, double, character and
    complex) of NAs are distinguished, and "S_compatible" to
    suppress the use of R-specific features such as 123L and to
    deparse integer values of a double vector with a trailing
    decimal point.

    The 'keepInteger' option now uses the suffix 'L' rather than
    as.integer() where possible (unless all entries are NA or
    "S_compatible" is also set).

    Other deparse options can now be added to "all" (which has not
    for some time actually switched on all options).

    Integer sequences m:n are now deparsed in that form.

    o   deparse() and dput() now include "keepInteger" and "keepNA" in
    their defaults for the 'control' argument.

    o   detach() now takes another argument, unload, which indicates
        whether or not to unload the package and then only cleans
        up the S4 methods if the package successfully unloads.

    o   There are new constants NA_integer_, NA_real_, NA_complex_ and
    NA_character_ to denote NAs of those types, and they will be
    used in deparsing in place of as.integer(NA) etc unless
    .deparseOpts() includes "S_compatible".

    o   dev.print() now recognizes 'screen devices' as all those with
    an enabled display list, rather than a hard-coded set.

    o   Objects of class "difftime" are now handled more flexibly. The
    units of such objects can now be accessed via a units()
    function, which also has a replacement form, and there are
    conversion methods to and from numeric, which also allow the
    specification of units.  Objects of this class can also be
    stored in data frames now.  A format() method has been added,
    and the print method was revised.

    o   New function environmentName() to give the print name of
    environments such as "namespace:base".
    This is now used by str().

    o   New function env.profile() provides R level access to summary
    statistics on environments.  In a related patch, new.env()
    now allows the user to specify an initial size for a hashed
    environment.

    o   file() can read the X11 clipboard selection as "X11_clipboard"
    on suitable X11-using systems.

    o   file("stdin") is now recognized, and refers to the process's
    'stdin' file stream whereas stdin() refers to the console.
    These may differ, for example for a GUI console, an embedded
    application of R or if --file= has been used.

    o   file_test() is now also available in package utils.
    (It is now private in package tools.)

    o   file.show() gains an 'encoding' argument.

    o   New functions formatUL() and formatOL() in package utils for
    formatting unordered (itemize) and ordered (enumerate) lists.

    o   The statistics reported when gcinfo(TRUE) are now of the amounts
    used (in Mb) and not of the amounts free (which are not really
    relevant when there are no hard limits, only gc trigger points).

    o   New function get_all_vars() to retrieve all the
    (untransformed) variables that the default method of
    model.frame() would use to create the model frame.

    o   interaction() has a new argument 'lex.order'.

    o   initialize() (in methods) now tries to be smarter about
        updating the new instance in place, thereby reducing
        copying.

    o   install.packages(dependencies = NA) is a new default, which is
    to install essential dependencies when installing from
    repositories to a single library.  As a result of this change,
    update.packages() will install any new dependencies of the
    packages it is updating (alongside the package in the same
    library tree).

    If 'lib' is not specified or is specified of length one and
    the chosen location is not a writable directory,
    install.packages() offers to create a personal library
    directory for you if one does not already exist, and to
    install there.

    o   is.atomic, is.call, is.character, is.complex, is.double (== is.real),
    is.environment, is.expression, is.function, is.integer,
    is.list, is.logical, is.null, is.object, is.pairlist,
    is.recursive, is.single and is.symbol (== is.name) are no
    longer internally S3 generic, nor can S4 methods be written
    for them.

    The "factor" methods of is.integer and is.numeric have been
    replaced by internal code.

    o   Added is.raw() for completeness.

    o   l10n_info() also reports if the current locale is Latin-1.

    o   levels<-(), names() and names<-() now dispatch internally for
    efficiency and so no longer have S3 default methods.

    o   .libPaths() now does both tilde and glob expansion.

    o   Functions lm(), glm() loess(), xtabs() and the default method
    of model.frame() coerce their 'formula' argument (if supplied)
    to a formula.

    o   max(), min() and range() now work with character vectors.

    o   message() has a new argument 'appendLF' to handle messages
    with and without newlines.  There is a new message class
    packageStartupMessage() that can be suppressed separately.

    o   A new function, method.skeleton() writes a skeleton version of
        a call to setMethod() to a file, with correct arguments and
        format, given the name of the function and the method signature.

    o   mode<- and storage.mode<- do slightly less copying.

    o   nls.control(* , printEval = FALSE, warnOnly = FALSE) are two new
    options to help better analyze (non-)convergence of nls(),
    thanks to Kate Mullen.

    nls() and summary(nls()) now contain more information and also
    print information about convergence.

    o   options(device = ) now accepts a function object as well as
    the name of a function.

    o   pdf() supports new values for 'paper' of "US" (same as
    "letter"), "a4r" and "USr" (the latter two meaning rotated to
    landscape).  postscript() also accepts paper = "US".

    o   persp() now respects the graphical pars 'cex.axis', 'cex.lab',
    'font.axis' and 'font.lab'.

    o   New faster internal functions pmax.int() and pmin.int() for
    inputs which are atomic vectors without classes (called by
    pmax/pmin where applicable).

    pmin/pmax are now more likely to work with classed objects:
    they work with POSIXlt datetimes, for example.

    o   postscript() now by default writes grey colors (including black
    and white) via 'setgray', which gives more widely acceptable
    output.  There are options to write pure RGB, CMYK or gray via
    the new argument 'colormodel'.

    o   rbind.data.frame() now ignores all zero-row inputs, as well as
    zero-column inputs (which it used to do, undocumented).  This
    is because read.table() can create zero-row data frames with
    NULL columns, and those cannot be extended.

    o   readChar() and writeChar() can now work with a raw vector.

    o   read.table(), write.table() and allies have been moved to package
    utils.

    o   rgb() now accepts the red, green and blue components in a
    single matrix or data frame.

    o   New utility function RShowDoc() in package 'utils' to find and
    display manuals and other documentation files.

    o   New .row_names_info() utility function finds the number of rows
    efficiently for data frames; consequently, dim.data.frame()
    has become very fast for large data frames with 'automatic'
    row names.

    o   RSiteSearch() now also allows to search postings of the 'R-devel'
    mailing list.

    o   screeplot() is now (S3) generic with a default method, thanks
    to a patch from Gavin Simpson.

    o   Experimental 'verbose' argument for selectMethod().  Might be
    replaced later by a better interface for method selection
    inspection.

    o   Added links to source files to the parsing routines, so that
    source() can now echo the original source and comments (rather
    than deparsing).  This affects example() and Sweave() as well.

    o   stack() and unstack() have been moved to package utils.

    o   strptime() now sets the "tzone" attribute on the result if
    tz != "".

    o   str.default() typically prints fewer entries of logical vectors.

    o   The RweaveLatex driver for Sweave() now supports two new
    options: expand=FALSE, to show chunk references in the output,
    and concordance=TRUE, to output the concordance between input
    and output lines.

    o   system() now takes the same set of arguments on all platforms,
    with those which are not applicable being ignored with a
    warning.  Unix-alikes gain 'input' and 'wait', and Windows
    gains 'ignore.stderr'.

    o   system.time() and proc.time() now return an object of class
    "proc_time" with a print() method that returns a POSIX-like
    format with names.

    o   Sys.getenv() has a new argument 'unset' to allow unset and set
    to "" to be distinguished (if the OS does).  The results of
    Sys.getenv() are now sorted (by name).

    o   New function Sys.glob(), a wrapper for the POSIX.2 function
    glob(3) to do wildcard expansion (on systems which have it,
    plus an emulation on Windows).

    o   Sys.setenv() is a new (and preferred) synonym for Sys.putenv().
    The internal C code uses the POSIX-preferred 'setenv' rather
    than 'putenv' where the former is available.

    o   New function Sys.unsetenv() to remove environment variables
    (on systems where unsetenv is implemented or putenv can remove
    variables, such as on Windows).

    o   text(), mtext(), strheight(), strwidth(), legend(), axis(),
    title(), pie(), grid.text() and textGrob() all attempt to
    coerce non-language annotation objects (in the sense of is.object)
    to character vectors.  This is principally intended to cover
    factors and POSIXt and Date objects, and is done via the
    new utility function as.graphicsAnnot() in package grDevices.

    o   tcltk:tk_select.list() now chooses the width to fit the widest item.

    o   {re,un}tracemem() are now primitives for efficiency and so
    migrate from 'utils' to 'base'.

    o   union(), interect(), setdiff() and setequal() now coerce their
    arguments to be vectors (and they were documented only to
    apply to vectors).

    o   uniroot() now works if the zero occurs at one of the ends of
    the interval (suggestion of Tamas Papp).

    o   There is a new function View() for viewing matrix-like
    objects in a spreadsheet, which can be left up whilst R is
    running.

    o   New function withVisible() allows R level access to the
    visibility flag.

    o   zip.file.extract() has been moved to package utils.


    o   A few more cases of subassignment work, e.g. <raw>[] <- <list>
    and <vector>[] <- <expression>, with suitable coercion of the LHS.

    o   There is a warning if \ is used unnecessarily in a string when
    being parsed, e.g. "\." where probably "\\." was intended.
    ("\." is valid, but the same as ".".)  Thanks to Bill Dunlap
    for the suggestion.

    o   Introduced the suffix L for integer literals to create
    integer rather than numeric values, e.g. 100L, 0x10L, 1e2L.

    o   Set the parser to give verbose error messages in case of
    syntax errors.

    o   The class "LinearMethodsList" has been extended and will be
    used to create list versions of methods, derived from the
    methods tables (environments).  The older recursive
    "MethodsList" class will be deprecated (by the release of
    2.5.0 if possible).

    o   There are more flexible ways to specify the default library
    search path.  In addition to R_LIBS and .Library, there are
    .Library.site (defaults to R_HOME/site-library) and
    R_LIBS_USER (defaults to a platform- and version-specific
    directory in ~/R).  See ?.libPaths for details.

    o   LAPACK has been updated to version 3.1.0.  This should cause
    only small changes to the output, but do remember that the sign
    of eigenvectors (and principal components) is indeterminate.

    o   PCRE has been updated to version 7.0.

    o   Several functions handle row names more efficiently:

    - read.table() and read.DIF() make use of integer row names
      where appropriate, and avoid at least one copy in assigning
      them.

    - data.frame() and the standard as.data.frame() methods avoid
      generating long dummy row names and then discarding them.

    - expand.grid() and merge() generate compact 'automatic' row
      names.

    - data.matrix() and as.matrix.data.frame() have a new argument
      'rownames.force' that by default drops 'automatic' row names.

    o   <data frame>[i, j] is substantially more memory-efficient when
    only a small part of the data frame is selected, especially
    when (part of) a single column is selected.

    o   Command-line R (and Rterm.exe under Windows) accepts the options
    '-f filename', '--file=filename' and '-e expression' to follow
    other script interpreters.  These imply --no-save unless
    --save is specified.

    o   Invalid bytes in character strings in an MBCS now deparse/print
    in the form "\xc1" rather than "<c1>", which means they can be
    parsed/scanned.

    o   Printing functions (without source attributes) and expressions
    now preserves integers (using the L suffix) and NAs (using
    NA_real_ etc where necessary).

    o   The 'internal' objects .helpForCall, .tryHelp and topicName are
    no longer exported from 'utils'.

    o   The internal regex code has been upgraded to glibc 2.5 (from 2.3.6).

    o   Text help now attempts to display files which have an \encoding
    section in the specified encoding via file.show().

    o   R now attempts to keep track of character strings which are
    known to be in Latin-1 or UTF-8 and print or plot them
    appropriately in other locales.  This is primarily intended
    to make it possible to use data in Western European languages
    in both Latin-1 and UTF-8 locales.  Currently scan(),
    read.table(), readLines(), parse() and source() allow
    encodings to be declared, and console input in suitable
    locales is also recognized.

    New function Encoding() can read or set the declared encodings
    for a character vector.

    o   There have been numerous performance improvements to the data
    editor on both Windows and X11.  In particular, resizing the
    window works much better on X11.

    o   Packages graphics and grid no longer require grDevices, as
    they might be used only with third-party devices.


DEPRECATED & DEFUNCT

    o   symbol.C() and symbol.For() are defunct, and have been replaced
    by wrappers that give a warning.

    o   Calling a builtin function with an empty argument is now always
    an error.

    o   The autoloading of ts() is defunct.

    o   The undocumented reserved word GLOBAL.ENV has been removed.
    (It was yet another way to get the value of the symbol
    .GlobalEnv.)

    o   The deprecated behaviour of structure() in adding a class when
    specifying with "tsp" or "levels" attributes is now defunct.

    o   unix() is now finally defunct, having been deprecated for
    at least seven years.

    o   Sys.putenv() is now deprecated in favour of Sys.setenv(),
    following the POSIX recommendation.

    o   Building R with --without-iconv is deprecated.

    o   Using $ on an atomic vector is deprecated (it was previously
    valid and documented to return NULL).

    o   The use of storage.mode<- for other than standard types (and
        in particular for value "single") is deprecated: use mode<-
    instead.


INSTALLATION

    o   A suitable iconv (e.g. from glibc or GNU libiconv) is
    required.  For 2.5.x only you can build R without it by
    configuring using --without-iconv.

    o   There is support again for building on AIX (tested on 5.2 and 5.3)
    thanks to Ei-ji Nakama.

    o   Autoconf 2.60 or later is used to create 'configure'.  This
    makes a number of small changes, and incorporates the changes
    to the detection of a C99-compliant C compiler backported for
    2.4.1.

    o   Detection of a Java development environment was added such that
    packages don't need to provide their own Java detection. Newly
    added make variables are JAVAC, JAVAH, JAR and JAVA_CPPFLAGS.

    R CMD javareconf was updated to look for the corresponding
    Java tools as well.

    In addition, Java detection honors user-supplied environment
    variables JAVA_CPPFLAGS, JAVA_LIBS and JAVA_LD_LIBRARY_PATH.

    o   Added workaround for reported non-POSIX sh on OSF1.  (PR#9375)

    o   'make install-strip' now works, stripping the executables and
    also the shared libraries and modules on platforms where
    'libtool' knows how to do so.

    o   Building R as a shared library and standalone nmath now
    installs pkg-config files 'libR.pc' and 'libRmath.pc' respectively.

    o   Added test for insufficiently complete implementation of sigaction.


C-LEVEL FACILITIES

    o   Functions str2type, type2char and type2str are now available in
    Rinternals.h.

    o   Added support for Objective C in R and packages (if available).

    o   R_ParseVector() has a new 4th argument 'SEXP srcfile' allowing
    source references to be attached to the returned expression list.

    o   Added ptr_R_WriteConsoleEx callback which allows consoles to
    distinguish between regular output and errors/warnings.
    To ensure backward compatibility it is only used if
    ptr_R_WriteConsole is set to NULL.


UTILITIES

    o   Additional Sweave() internal functions are exported to help
    writing new drivers, and RweaveLatexRuncode() is now created
    using a helper function (all from a patch submitted by Seth
    Falcon).

    o   The following additional flags are accessible from R CMD config:
    OBJC, OBJCFLAGS, JAR, JAVA, JAVAC, JAVAH, JAVA_HOME, JAVA_LIBS
    and JAVA_CPPFLAGS.

    o   R CMD build now takes the package name from the DESCRIPTION
    file and not from the directory.  (PR#9266)

    o   checkS3methods() (and hence R CMD check) now checks agreement
    with primitive internal generics, and checks for additional
    arguments in methods where the generic does not have a '...'
    argument.

    codoc() now knows the argument lists of primitive functions.

    o   R CMD INSTALL and R CMD REMOVE now use as the default library
    (if -l is not specified) the first library that would be used
    if R were run in the current environment (and they run R to
    find it).

    o   There is a new front-end Rscript which can be used for #!
    scripts and similar tasks.  See help("Rscript") and 'An
    Introduction to R' for further details.

    o   R CMD BATCH (not Windows) no longer prepends
    'invisible(options(echo = TRUE))' to the input script.  This
    was the default unless --slave is specified and the latter is
    no longer overridden.

    On all OSes it makes use of the -f argument to R, so
    file("stdin") can be used from BATCH scripts.

    On all OSes it reports proc.time() at the end of the script
    unless q() is called with options to inhibit this.

    o   R CMD INSTALL now prepends the installation directory (if
    specified) to the library search path.

    o   Package installation now re-encodes R files and the NAMESPACE
    file if the DESCRIPTION file specifies an encoding, and sets
    the encoding used for reading files in preparing for
    LazyData.  This will help if a package needs to be used in
    (say) both latin1 and UTF-8 locales on different systems.

    o   R CMD check now reports on non-ASCII strings in datasets.
    (These are a portability issue, which can be alleviated by
    marking their encoding: see 'Writing R Extensions'.)

    o   Rdiff now converts CRLF endings in the target file, and
    converts UTF-8 single quotes in either to ASCII quotes.

    o   New recommended package 'codetools' by Luke Tierney provides
    code-analysis tools.  This can optionally be used by 'R CMD
    check' to detect problems, especially symbols which are not
    visible.

    o   R CMD config now knows about LIBnn .

    o   New recommended package 'rcompgen' by Deepayan Sarkar
    provides support for command-line completion under the Unix
    terminal interface (provided readline is enabled) and the
    Windows Rgui and Rterm front ends.


BUG FIXES

    o   gc() can now report quantities of 'Vcells' in excess of 16Gb on
    64-bit systems (rather than reporting NA).

    o   Assigning class "factor" to an object now requires it has
    integer (and not say double) codes.

    o   structure() ensures that objects with added class "factor" have
    integer codes.

    o   The "formula" and "outer" attributes of datasets 'ChickWeight',
    'CO2', 'DNase', 'Indometh', 'Loblolly', 'Orange' and 'Theoph'
    now have an empty environment and not the environment used to
    dump the datasets in the package.

    o   Dataset 'Seatbelts' now correctly has class c("mts", "ts").

    o   str() now labels classes on data frames more coherently.

    o   Several 'special' primitives and .Internals could return
    invisibly if the evaluation of an argument led to the
    visibility flag being turned off.  These included
    as.character(), as.vector(), call(), dim(), dimnames(),
    lapply(), rep(), seq() and seq_along().  Others (e.g. dput()
    and print.default()) could return visibly when this was not
    intended.

    o   Several primitives such as dim() were not checking the number
    of arguments supplied before method dispatch.

    o   Tracing of primitive functions has been corrected.  It should
    now be the case that tracing either works or is not allowed
    for all primitive functions. (Problems remain if you make a
    primitive into a generic when it is being traced.  To be fixed
    later.)

    o   max.col() now omits infinite values in determining the
    relative tolerance.

    o   R CMD Sweave and R CMD Stangle now respond to --help and --version
    like other utilities.

    o   .libPaths() adds only existing directories (as it was
    documented to, but could add non-directories).

    o   setIs() and setClassUnion() failed to find some existing
    subclasses and produced spurious warnings, now fixed.

    o   data.frame() ignored 'row.names' for 0-column data frames, and
    no longer treats an explicit row.names=NULL differently from
    the default value.

    o   identical() looked at the internal structure of the
    'row.names' attribute, and not the value visible at R level.

    o   abline(reg) now also correctly works with intercept-only lm
    models, and abline() warns more when it's called illogically.

    o   warning() was truncating messages at
    getOption("warning.length") - 1 (not as documented), with no
    indication.  It now appends '[... truncated]'.

    o   Stangle/Sweave were throwing spurious warnings if options
    'result' or 'strip.white' were unset.

    o   all.equal() was ignoring 'check.attributes' for list and
    expression targets, and checking only attributes on raw vectors.
    Logical vectors were being compared as if they were numeric,
    (with a mean difference being quoted).

    o   Calculating the number of significant digits in a number was
    itself subject to rounding errors for digits >= 16.  The
    calculation has been changed to err on the side of slightly
    too few significant digits (but still at least 15) rather than
    far too many.  (An example is print(1.001, digits=16).)

    o   unlink() on Unix-alikes failed for paths containing spaces.

    o   substr() and friends treated NA 'start' or 'stop' incorrectly.

    o   merge(x, y, all.y = TRUE) would sometimes incorrectly return
    logical columns for columns only in y when there were no
    common rows.

    o   read.table(fn, col.names=) on an empty file returned NULL
    columns, rather than logical(0) columns (which is what results
    from reading a file with just a header).

    o   grid.[xy]axis(label=logical(0)) failed.

    o   expression() was unnecessarily duplicating arguments.

    o   as.expression(<list>) returned a single-element expression
    vector, which was not compatible with S: it now copies lists
    element-by-element.

    o   supsmu(periodic = TRUE) could segfault.  (PR#9502, detection
    and patch by Bill Dunlap.)

    o   pmax/pmin called with only logical arguments did not coerce to
    numeric, although they were documented to do so (as max/min
    do).

    o   methods() did not know that cbind() and rbind() are internally
    generic.

    o   dim(x) <- NULL removed the names of x, but this was always
    undocumented.  It is not clear that it is desirable but it is
    S-compatible and relied on, so is now documented.

    o   which(x, arr.ind = TRUE) did not return a matrix (as
    documented) if 'x' was an array of length 0.

    o   C-level duplicate() truncated CHARSXPs with embedded nuls.

    o   Partial matching of attributes was not working as documented
    in some cases if there were more than two partial matches or
    if "names" was involved.

    o   data(package=character(0)) was not looking in ./data as
    documented.

    o   summary.mlm() failed if some response names were "" (as can
    easily happen if cbind() is used).

    o   The postscript() and pdf() drivers shared an encoding list
    but used slightly different formats.  This caused problems if
    both were used with the same non-default encoding in the same
    session.  (PR#9517)

    o   The data editor was not allowing Inf, NA and NaN to be entered
    in numerical columns.  It was intended to differentiate
    between empty cells and NAs, but did not do so: it now does
    so for strings.

    o   supsmu() could segfault if all cases had non-finite values.
    (PR#9519)

    o   plnorm(x, lower.tail=FALSE) was returning the wrong tail for
    x <= 0.  (PR#9520)

    o   which.min() would not report a minimum of +Inf, and
    analogously for which.max().  (PR#9522)

    o   'R CMD check' could fail with an unhelpful error when checking
    Rd files for errors if there was only one file and that had a
    serious error.  (PR#9459)

    o   try() has been reimplemented using tryCatch() to solve two
    problems with the original implementation: (i) try() would run
    non-NULL options("error") expressions for errors within a try, and
    (ii) try() would catch user interrupts.

    o   str(obj) could fail when obj contained a dendrogram.

    o   Using <data frame>[, <last column>] <- NULL failed (PR#9565)

    o   choose(n, k) could return non-integer values for integer n and
    small k on some platforms.

    o   nclass.scott(x) and nclass.FD(x) no longer return NaN when var(x)
    or IQR(x) (respectively) is zero.

    hist() now allows breaks = 1 (which the above patch will
    return), but not breaks = Inf (which gave an obscure error).

    o   strptime("%j") now also works for the first days of Feb-Dec.
    (PR#9577)

    o   write.table() now recovers better if 'file' is an unopened
    connection.  (It used to open it for both the column names and
    the data.)

    o   Fixed bug in mosaicplot(sort=) introduced by undocumented
    change in R 2.4.1 (changeset r39655).

    o   contr.treatment(n=0) failed with a spurious error message.
    (It remains an error.)

    o   as.numeric() was incorrectly documented: it is identical to
    as.double.

    o   jitter(rep(-1, 3)) gave NaNs. (PR#9580)

    o   max.col() was not random for a row of zeroes. (PR#9542)

    o   ansari.test(conf.int=TRUE, exact=FALSE) failed.

    o   trace() now works on S3 registered methods, by modifying the
    version in the S3 methods table.

    o   rep(length=1, each=0) segfaulted.

    o   postscript() could overflow a buffer if used with a long
    'command' argument.

    o   The internal computations to copy complete attribute lists did
    not copy the flag marking S4 objects, so the copies no longer
    behaved like S4 objects.

    o   The C code of nlminb() was altering a variable without
    duplicating it.  (This did not affect nlminb() but would have
    if the code was called from a different wrapper.)

    o   smooth(kind = "3RS3R") (the current default) used .C(DUP =
    FALSE) but altered its input argument.  (This was masked by
    duplication in as.double.)

    o   The signature for the predefined S4 method for as.character()
    was missing '...' .

    o   readBin(<raw vector>) could read beyond the end of the vector
    when size-changing was involved.

    o   The C entry point PrintValue (designed to emulate auto-printing)
    would not find show() for use on S4 objects, and did not have
    the same search path (for show(), print() and print() methods)
    as auto-printing.  Also, auto-printing and print() of S4
    objects would fail to find 'show' if the methods name space was
    loaded but the package was not attached (or otherwise not in
    the search path).

    o   print() (and auto-printing) now recognize S4 objects even when
    'methods' is not loaded, and print a short summary rather than
    dump the internal structure.

    o   Sweave and Stangle had problems due to partial matching of code
    chunk names when run with split=TRUE.

    o   install.packages() on a source package now ensures that
    R CMD INSTALL sees the same library search path as
    install.packages() did when computing dependencies.

    o   density() now ensures its 'y' values are non-negative. (PR#8876)

    o   is.finite() and is.infinite() (and many other primitives)
    are not internally generic and so do not support S4 methods,
    which can no longer be set. (PR#7951)

    o   nls(algorithm = "port") now accepts a list 'start' argument,
    as for the other methods (and as documented).

    o   Standard errors from the "ar" method of predict() could be
    wrong for the last p predictions for models near
    non-stationarity. (PR#9614)




    **************************************************
    *                        *
    *          2.4 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 2.4.1 patched

NEW FEATURES

    o   The Simplified Chinese translations have been completed.


BUG FIXES

    o   co.intervals() sometimes failed to cover the largest value.

    o   tempfile() is now random across sessions as well as within a
    session.  (On some systems it would give the same hex suffix
    at the start of each session.)

    o   Added infinite recursion test to internal function isMissing.
    (PR#9426)

    o   The "Date" and "POSIXt" methods for cut() were not choosing
    the first day of the year for breaks = "years".
    (In part, PR#9433.)

    o   R is now able to deparse/print invalid multibyte strings in
    MBCS locales (such as UTF-8) using hex escapes.  This means
    that e.g. demo(Hershey) works in all such locales.

    o   optimize() could give incorrect answers in some rare problems
    with exact symmetry about the midpoint of the interval
    supplied.  (PR#9438)

    o   The residuals from an lm() fit with no coefficients but an
    offset were incorrect.

    o   oneway.test() was expecting a literal formula and did not
    accept a variable containing a formula.

    o   The legacy Quartz device (used by console R) displayed its
    window outside the screen estate in some dual-head setups.
    Now it will be always displayed in the center of the main
    screen.

    o   read.ftable() was not functional on non-seekable connections
    such as URLs.

    o   Some large memory allocations could cause segfaults or
    crashes (e.g. followup to PR#9557).

    o   Sweave() would drop characters from the end of chunk names
        ending in "R".  (PR#9567)

    o   library(), i.e. its internal checkConflicts(), now (again) prints
    "The following object(s) are masked .." only once per masked package.

    o   methods:::cbind(x) {one argument} now works, calling cbind2(x)
    when 'x' is an S4 object.



        CHANGES IN R VERSION 2.4.1


INSTALLATION

    o   The extraction of info from Subversion for an SVN checkout now
    also works for svn >= 1.4.0.  However, on Windows the 'Last
    Changed Date' will be in the local timezone, and not in GMT as
    previously.

    o   configure uses code borrowed from autoconf 2.60 to try harder to
    ensure that a C99-compliant compiler is used.  (It does so by
    appending to CC.)  This avoids problems with systems such as FC5
    which override CFLAGS and thereby lose flags such as -std=gnu99.


NEW FEATURES

    o   rainbow(), heat.colors(), terrain.colors(), topo.colors() and
    cm.colors() all gain an 'alpha' argument to be passed to hsv().

    o   dput() will give an incorrect representation of the row names
    of a data frame with integer row names.  This is now corrected
    when the object is recreated.


C-LEVEL FACILITIES

    o   Using STRICT_R_HEADERS applies to more reported clashes with
    Windows headers, including Calloc and Realloc.  These and
    Free need to be prefixed by R_ when STRICT_R_HEADERS is defined.


DEPRECATED & DEFUNCT

    o   The previously undocumented behaviour of structure() in adding a
    class when specifying "tsp" or "levels" attributes is now
    deprecated (with a warning).


BUG FIXES

    o   Fixed warning() to use .dfltWarn intead of .dfltStop for default
    handling (PR#9274).

    o   R would slow down when the product of the length of a vector and
    the length of a character vector used to subset it exceeded 2^31.
    (PR#9280)

    o   merge() now allows zero-row data frames.

    o   add1.lm() had been broken by other changes for weighted fits.

    o   axis.POSIXct() would sometimes give the wrong labels.

    o   Help for a method call would fail. (PR#9291)

    o   gzfile() returned an object of class "file" not "gzfile". (PR#9271)

    o   load()ing from a connection had a logic bug in when it closed
    the connection. (PR#9271)

    o   The lowess() algorithm is unstable if the MAD of the residuals
    becomes (effectively) zero: R now terminates the iterations at
    that point.  (This may result in quite different answers.)

    The 'delta' argument was incorrectly documented.  (PR#9264)

    o   abbreviate() would only work for strings of up to 8191 bytes,
    but this was not checked.  Now longer strings are errors.

    o   Drawing X11 rotated text was buggy for VERY small (negative)
    angle of rotation.  Reported by Ben Bolker. (PR#9301)

    o   The X11 data editor would crash in an MBCS locale if R was
    compiled with FC's CFLAGS that add buffer overflow and
    stack-smashing detection.

    o   rect() was not accepting border=NA in some cases involving
    cross-hatching.

    o   Fixes to S4 group generics to ensure that the correct number of
    active arguments are in the signature of the group and all
    members.  Also a fix to keep the 'groupMembers' slot up to date.

    o   S4 group generic "Logic" (with '&', '|', but not '!') has been
    created, following the green book (apart from '!').

    o   removeClass() now takes care to remove any subclass references
    to the deleted class.

    o   mle() (in stats4) might not have worked as intended when the
    order of parameters in 'start' differed from that in the
    log-likelihood.  (PR#9313)

    o   dotchart() now properly restores par() settings after itself.

    o   system() on Mac OS X was blocking arbitrary signals during the
    call although only SIGPROF was meant to be blocked.

    o   methods cached via callNextMethod() and (sometimes) as() were
    being cached as directly specified although in fact they were
    inherited.  Caused problems in later search for inherited methods.

    o   str() works properly for method definitions and other S4-classed
    function objects.

    o   JAVA_LIBS are now set correctly on MacOS X.

    o   Fix null-termination issue suspected of causing crash with Fedora
    Extra RPMS (PR#9339, Justin Harrington, analysis and fix from Bill
    Dunlap).

    o   Name spaces restored via a saved session silently failed to cache
    their methods because the methods package was not yet
    attached.  Fixed by attaching methods before restoring data.

    o   rbind()ing a list to a data frame generated invalid row names,
    which were an error in 2.4.0. (PR#9346)

    o   boxplot.stats(x) now returns the correct minimum instead of an
    error for x <- c(1,Inf,Inf,Inf), and hence boxplot(x) "works".

    o   promptClass() now uses \linkS4class{<ClassName>} instead of
    of \link{<ClassName>-class}.

    o   gc() no longer reports nonsense values for the number of used
    Vcells if the true value exceeds 2^31 (and hence over 16Gb of
    heap is in use): it now reports NA. (PR#9345)

    o   rapply() now detects more user errors in supplying arguments. (PR#9349)

    o   boxplot() was ignoring argument 'boxfill'. (PR#9352)

    o   plot.lm(which = 6, id.n = 0) did not work. (PR#9333)

    o   .deparseOpts("delayPromises") was not matching the C code,
    returning 64 rather than 32.

    o   bxp() could use partial matching on 'pars' when finding
    defaults for some of its parameters, e.g. a setting of 'cex.axis'
    in 'pars' or inline was used to set a default for 'outcex'.

    o   acf() now allows lag.max = 0 except when type="partial", and
    forces the lag 0 autocorrelation to 1. (PR#9360)

    o   hist(*, include.lowest=., right=., plot=FALSE) does not warn
    anymore, (PR#9356) and more.

    o   Some bugs in caching superclass/subclass relations and in
    removing those relations on detach and on removeClass() have
    been fixed.

    o   readBin() could return one too many strings if 'n' was an
    over-estimate. (PR#9361)

    o   A request for an opaque colour in the pdf() device after a
    translucent one did not set the transparency back to opaque in
    2.4.0.

    Semi-transparent background colours were not being plotted on
    the pdf() device.

    o   plot.lm(which=5) in the case of constant leverage re-ordered
    the factor levels but not the residuals, so the labelling by
    factor level was often incorrect.

    o   packBits() was not accepting a logical argument. (PR#9374)

    o   make install was omitting doc/FAQ and doc/RESOURCES.

    o   A two-sample t.test(x, y, var.equal=TRUE) did not allow one of the
    groups to be of size one.

    o   The "ts" method for print() failed on some corrupted objects
    of class "ts", e.g. those without a "tsp" attribute.

    o   structure() reordered the "class" value given if there was a
    "tsp" value specified.

    o   pairs() now does pass appropriate parts of '...' to the
    'diag.panel' argument. (PR#9384)

    o   plot.lm() was using an incorrect estimate of dispersion for
    some GLMs (including family=binomial and family=poisson).
    (PR#9316)

    o   Subsetting operators were setting R_Visible too early, so
    assignments in arguments could make the result invisible.
    (PR#9263)

    o   The tk-GUI was displaying a warning due to an extra comma in
        the list of manuals (PR#9396)

    o   packageDescription() now gives an explicit error on a corrupt
    DESCRIPTION file.

    o   There was a scoping issue with tcltk callbacks given as
        unevaluated expressions. This has only been partially fixed, a
        complete fix probably requires redesign.

    o   trace() had its return value documented incorrectly and was
    sometimes visible when it should not have been.

    o   pchisq() would sometimes use the wrong tail when calculating
    non-central probabilities with lower.tail = FALSE. (PR#9406)

    o   rm() could remove the wrong objects when passed an expression.
    (PR#9399)  Now only names are allowed in the '...' argument,
    and the incorrect documentation of what happened with
    character objects is corrected.

    o   url() was not supporting 'encoding' except on file:// URLs.



        CHANGES IN R VERSION 2.4.0


USER-VISIBLE CHANGES

    o   The startup message now prints first the version string and
    then the copyright notice (to be more similar to R --version).

    o   save() by default evaluates promise objects.  The old behaviour
    (to save the promise and its evaluation environment) can be
    obtained by setting the new argument 'eval.promises' to FALSE.
    (Note that this does not apply to promises embedded in
    objects, only to top-level objects.)

    o   The functions read.csv(), read.csv2(), read.delim(),
    read.delim2() now default their 'comment.char' argument to "".
    (These functions are designed to read files produced by other
    software, which might use the # character inside fields, but
    are unlikely to use it for comments.)

    o   The bindings in the base environment/name space (currently the
    same thing) are now locked.  This means that the values of
    base functions cannot be changed except via
    assignInNamespace() and similar tricks.

    o   [[ on a factor now returns a one-element factor (and not an
    integer), as.list() on a factor returns a list of one-element
    factors (and not of character vectors), and unlist() on a list
    of factors returns a factor (and not an integer vector).
    These changes may affect the results of sapply() and lapply()
    applied to factors.

    o   mauchly.test() now returns the W statistic (for comparability
    with SAS and SPSS), rather than the z (which was accidentally
    not named in the output)

    o   sort(x, decreasing = FALSE, ...) is now a generic function.
    This means that 'partial' is no longer the second argument,
    and calls which used positional matching may be incorrect: we
    try to detect them.

    o   See the section on 'Changes to S4 methods': all packages
    depending on 'methods' need to be re-installed.


NEW FEATURES

    o   agrep(), grep(), strwrap(), strtrim(), substr() and
    related functions now coerce arguments which should be
    character via as.character() rather than internally (so method
    dispatch takes place, e.g. for factors).

    chartr(), charfold(), tolower() and toupper() now coerce their
    main argument if necessary to a character vector via
    as.character().

    Functions which work element-by-element on character vectors
    to give a character result now preserve attributes including
    names, dims and dimnames (as suggested by the Blue Book
    p. 144).  Such functions include charfold(), chartr(), gsub(),
    strtrim(), sub(), substr(), tolower() and toupper().  (Note
    that coercion of a non-character argument may lose the
    attributes.)

    agrep(value = TRUE) preserves names for compatibility with
    grep().

    nchar() has always preserved dims/dimnames (undocumented
    before) and now also preserves names.


    o   .Deprecated and .Defunct take a new parameter, msg, that allows
    for the specification of the message printed and facilitates
    deprecation of calling sequences etc.

    o   .Fortran() will map 'name' to lower case, and will work with
    'name' containing underscores.

    o   The default is now .saveRDS(compress = TRUE)

    o   The :: operator now also works for packages without name spaces
    that are on the search path.

    o   [[ on a list does not duplicate the extracted element unless
    necessary.  (It did not duplicate in other cases, e.g. a
    pairlist.)

    o   argsAnywhere() works like args() on non-exported functions.

    o   as.data.frame() gains a '...' argument.

    o   Added an as.data.frame() method for class "ftable".

    o   as.list(<an expression>) is now handled by internal code and
    no longer loses attributes such as names.

    as.list(<a list>) no longer duplicates (unnecessarily).

    o   as.POSIX[cl]t can now convert character strings containing
    fractional seconds.

    o   attach() can now attach a copy of an environment.

    o   available.packages() and installed.packages() gain a 'fields'
    argument thanks to Seth Falcon.

    o   axis.POSIXct() uses a different algorithm for ranges of 2 to 50
    days that will mark days at midnight in the current timezone
    (even if the graph crosses a DST change).

    o   body<-() and formals<-() default to envir = environment(fun),
    that is they do not by default change the environment.
    (Previously they changed it to parent.frame().)

    o   New function combn(x, m, ..) for computing on all combinations of
    size 'm' (for small 'm' !).

    o   The cumxxx() functions now handle logical/integer arguments
    separately from numeric ones, and so return an integer result
    where appropriate.

    o   data.frame() has a new argument 'stringsAsFactor'.  This and
    the default for read.table(as.is=) are set from the new global
    option 'stringsAsFactors' via the utility function
    default.stringsAsFactors().

    o   dev.interactive() now has an optional argument 'orNone'.

    o   df() now has a noncentrality argument 'ncp', based on a
    contribution by Peter Ruckdeschel.

    o   example() gains an argument 'ask' which defaults to "TRUE when
    sensible", but the default can be overridden by setting option
    'example.ask'.

    o   expand.grid() now has an argument 'KEEP.OUT.ATTRS' which can
    suppress (the potentially expensive) "out.attrs" attribute. It no
    longer returns an extraneous 'colnames' attribute.

    o   The subset and subassign methods for factors now handle factor
    matrices, and dim() can be set on a factor.

    o   There is now a format() method for class "ftable".

    o   head(x, n) and tail(x, n) now also work for negative arguments,
    thanks to Vincent Goulet.

    o   head.matrix() and tail.matrix() are no longer hidden, to be used
    for building head() and tail() methods for other classes.

    o   If help() finds multiple help files for a given topic, a menu
    of titles is used to allow interactive choice.

    o   help.search() now rebuilds the database if 'package' specifies a
    package not in the saved database.

    o   hist(*, plot = FALSE) now warns about unused arguments.

    o   history() gains a 'pattern' argument as suggested by Romain
    Francois.

    o   integer(0) now prints as that rather than "numeric(0)" (it
    always deparsed as "integer(0)").

    o   interaction(..., drop=TRUE) now gives the same result as
    interaction(...)[,drop=TRUE] (it used to sometimes give a
    different order for the levels).

    o   lag.plot() produces a conventional plot (not setting mfrow) if
    only one plot is to be produced.

    o   lapply() does much less copying.  Vector X are handled without
    duplication, and other types are coerced via as.list().  (As a
    result, package 'boot' runs its examples 4% faster.)

    lapply(<a pairlist>) now coerces to a list (rather than traverse
    the pairlist from the beginning for each item).

    o   legend() has new parameters 'box.lwd' and 'box.lty'.

    o   lines() gains a simple method for isoreg() results.

    o   load() no longer coerces pairlists to lists (which was
    undocumented, but has been happening since 1998).

    o   make.link() now returns an object of class "link-glm".
    The GLM families accept an object of this class for their
    'link' argument, which allows user-specified link functions.
    Also, quasi() allows user-specified variance functions.

    o   mapply() uses names more analogously to lapply(), e.g..

    o   matplot() now accepts a 'bg' argument similarly to plot.default() etc.

    o   median() is now generic, and its default method uses mean()
    rather than sum() and so is more widely applicable (e.g. to dates).

    o   Dummy functions memory.size() and memory.limit() are available
    on Unix-alikes, for people who have not noticed that documentation
    is Windows-specific.

    o   merge() works more efficiently when there are relatively few
    matches between the data frames (for example, for 1-1
    matching).  The order of the result is changed for 'sort = FALSE'.

    o   merge() now inserts row names as a character column and not a
    factor: this makes the default sort order more comprehensible.

    o   Raw, complex and character vectors are now allowed in model
    frames (there was a previously undocumented restriction to
    logical, integer and numeric types.).  Character vectors in a
    formula passed to model.matrix() are converted to factors and
    coded accordingly.

    o   modifyList() utility, typically for housekeeping nested lists.

    o   x <- 1:20; y <- rnorm(x); nls(y ~ A*exp(-x^2/sig)) no longer
    returns an unhelpful error message.  In this and similar cases, it
    now tries a wild guess for starting values.

    o   Ops.difftime() now handles unary minus and plus.

    o   Ops.Date() and Ops.POSIXt() now allow character arguments
    (which are coerced to the appropriate class before comparison,
    for Ops.POSIXt() using the current time zone).

    o   There is a new option(max.contour.segments = 25000) which can be
    raised to allow extremely complex contour lines in contour()
    and contourLines().  (PR#9205)

    o   options(max.print = N) where N defaults to 99999 now cuts printing
    of large objects after about N entries.  print(x, ..., max = N)
    does the same for the default method and those building on
    print.default().

    options("menu.graphics") controls if graphical menus should be
    used when available.

    options("par.ask.default") allows the default for par("ask")
    to be set for a newly-opened device.  (Defaults to FALSE, the
    previous behaviour.)

    The way option("papersize") is set has been changed.  On
    platforms which support the LC_PAPER locale category, the
    setting is taken first from the R_PAPERSIZE environment
    variable at run time, then from the LC_PAPER category
    ("letter" for _US and _CA locales and "a4" otherwise).  On
    other platforms (including Windows and older Unixen), the
    choice is unchanged.

    o   package.skeleton() gains arguments 'namespace' and
    'code_files'.

    o   par(ask=TRUE) now only applies to interactive R sessions.

    o   parse() now returns up to 'n' expressions, rather than fill
    the expressions vector with NULL.  (This is now compatible
    with S.)

    o   The 'version' argument for pdf() is now increased
    automatically (with a warning) if features which need a higher
    level are used.

    o   pie() now allows expressions for 'labels', and empty slices.

    o   There is a new '%.%' operator for mathematical annotations
    (plotmath) which draws a centred multiplication dot
    (a \cdot in LaTeX), thanks to Uwe Ligges.

    o   predict.lm() gains a 'pred.var' argument.  (Wishlist PR#8877.)

    o   print.summary.{aov,glm,lm,nls} and print.{aov,glm} make use of
    naprint() to report when na.action altered the model frame.

    o   print.table(T, zero.print=ch) now also replaces 0 by ch when T is
    non-integer with integer entries.

    o   Recursive rapply() which is similar to lapply but used
    recursively and can restrict the classes of elements to which
    it is applied.

    o   r2dtable() has been moved to package 'stats'.

    o   New function read.DIF() to read Data Interchange Format files,
    and (on Windows) this format from the clipboard.

    o   New experimental function readNEWS() to read R's own "NEWS" file
    and similarly formatted ones.

    o   readLines() has a new argument 'warn' to suppress warnings:
    the default behaviour is still to warn.

    o   reg.finalizer() has a new argument 'onexit' to parallel the
    C-level equivalent R_RegisterFinalizerEx.

    o   rep() is now a primitive function and under some conditions
    very much faster: rep.int() is still a little faster (but does
    less). (Because it is primitive there are minor changes to the
    call semantics: see the help page.)

    o   The 'row.names' of a data frame may be stored internally as an
    integer or character vector.  This can result in considerably
    more compact storage (and more logical row names from rbind)
    when the row.names are 1:nrow(x).  However, such data frames
    are not compatible with earlier versions of R: this can be
    ensured by supplying a character vector as 'row.names'.

    row.names() will always return a character vector, but direct
    access to the attribute may not.

    The internal storage of row.names = 1:n just records 'n', for
    efficiency with very long vectors.

    The "row.names" attribute must be a character or integer
    vector, and this is now enforced by the C code.

    o   The "data.frame" and "matrix" methods for rowsum() gain an 'na.rm'
    argument.

    o   Experimental support for memory-use profiling via Rprof(),
    summaryRprof(), Rprofmem() and tracemem().

    o   save.image() [also called by sys.save.image() and hence from q()]
    now defaults to saving compressed binary images.  To revert to
    the previous behaviour set option "save.image.defaults": see
    ?save.image.

    o   There is a new primitive seq.int() which is slightly more
    restricted than seq() but often very much faster, and new
    primitives seq_along() and seq_len() which are faster still.

    o   serialize(connection = NULL) now returns a raw vector (and not
    a character string).  unserialize() accepts both old and new
    formats (and has since 2.3.0).

    o   setwd() now returns the previously current directory (invisibly).

    o   The function sort() is now sort.int(), with a new generic
    function sort() which behaves in the same way (except for the
    order of its argument list) for objects without a class, and
    relies on the '[' method for objects with a class (unless a
    specific method has been written, as it has for class "POSIXlt").

    o   sort.list() now implements complex vectors (PR#9039), and how
    complex numbers are sorted is now documented.

    o   spline() and splinefun() now follow approx[fun] to have an argument
    'ties = mean' which makes them applicable also when 'x' has
    duplicated values.

    o   str(x) does not print the S3 "class" attribute when it is the
    same as 'mode' (which is printed anyway, possibly abbreviated)
    and it puts it beside mode for atomic objects such as S3 class
    "table".

    o   str(<data.frame>) now outputs 'data.frame' instead of
    `data.frame'; this may affect some strict (Package) tests.

    o   str() now takes also its defaults for 'vec.len' and 'digits.d'
    from options('str') which can be set by the new strOptions().

    o   symnum() has a new argument 'numeric.x' particularly useful for
    handling 0/1 data.

    o   Sys.getlocale() and Sys.setlocale() support LC_MESSAGES,
    LC_PAPER and LC_MEASUREMENT if the platform does.

    o   Sweave has a new options 'pdf.encoding' and 'pdf.version' for
    its Rweave driver.

    o   The character vector used by an output textConnection() has
    a locked binding whilst the connection is open.

    There is a new function textConnectionValue() to retrieve the
    value of an output textConnection().

    o   traceback() gains a 'max.lines' argument.
    .Traceback is no longer stored in the workspace.

    o   warning(immediate. = TRUE) now applies to
    getOption("warn") < 0 and not just == 0.

    o   warnings() is now an accessor function for 'last.warning' (which
    is no longer stored in the workspace) with a print() method.


    o   The internal internet download functions have some new
    features from libxml 2.6.26.

    o   There is an option "HTTPUserAgent" to set the User Agent in R
    download requests etc.  Patch from S. Falcon.

    o   PCRE has been updated to version 6.7.

    o   The C function substituteList now has tail recursion expanded
    out, so C stack overflow is less likely.  (PR#8141, fix by
    Kevin Hendricks)

    o   The (somewhat soft) 1023/4 byte limit on command lines is now
    documented in 'An Introduction to R'.

    o   The maximum number of open connections has been increased from
    50 to 128.

    o   There is a new manual 'R Internals' on R internal stuctures
    plus the former appendices of 'Writing R Extensions'.

    o   The autoloads introduced at the package re-organization have been
    almost completely removed: the one that remains is for ts().

    o   The setting of the various Java configuration variables has been
    improved to refer to JAVA_HOME, and they are now documented in
    the R-admin manual.

    o   It is (again) possible to calculate prediction intervals from
    "lm" objects for the original data frame, now with a warning
    that the intervals refer to future observations.  Weighted
    intervals have also been implemented, with user-specifiable
    weights. Warnings are given in cases where the default
    behaviour might differ from user expectations.  See the
    ?predict.lm for details.


CHANGES TO S4 METHODS

    o   The default prototype object for S4 classes will have its own
    internal type in 2.4.0, as opposed to being an empty list (the
    cause of several errors in the code up to 2.3.1).  Note that old
    binary objects, including class definitions, will be
    inconsistent with the type, and should be recreated.

    o   S4 method dispatch has been completely revised to use cached
    generic functions and to search for the best match among
    inherited methods.  See ?Methods and
    http://developer.r-project.org/howMethodsWork.pdf

    o   Objects created from an S4 class are now marked by an internal flag,
    tested by isS4() in R and by macro IS_S4_OBJECT() in C.  This
    is an efficient and reliable test, and should replace all
    earlier heuristic tests.

    o   Some changes have been made to automatic printing of S4
    objects, to make this correspond to a call to show(), as per
    'Programming with Data'.

    o   S4 generic and class definitions are now cached when the related
    package is loaded.  This should improve efficiency and also
    avoid anomalous situations in which a class or generic cannot
    be found.

    o   trace() now creates a new S4 class for the traced object if
    required.  This allows tracing of user-defined subclasses of
    "function".


DEPRECATED & DEFUNCT

    o   The re-named tcltk functions tkcmd, tkfile.tail, tkfile.dir,
    tkopen, tkclose, tkputs, tkread are now defunct.

    o   Argument 'col' of bxp() has been removed: use 'boxfill'.

    o   Use of NULL as an environment is now an error.

    o   postscriptFont() is defunct: use Type1Font().

    o   La.chol() and La.chol2inv() are defunct (they were the same as
    the default options of chol() and chol2inv).

    o   La.svd(method = "dgesvd") is defunct.

    o   Files install.R and R_PROFILE.R in packages are now ignored
    (with a warning).

    o   The following deprecated command-line options to INSTALL have
    been removed (use the fields in the DESCRIPTION file instead):
    -s --save --no-save --lazy --no-lazy --lazy-data --no-lazy-data

    o   Graphical parameter 'tmag' is obsolete.

    o   mauchley.test() (package 'stats') is now defunct.

    o   symbol.C() and symbol.For() are deprecated.  They are required in
    S for use with is.loaded(), but are not so required in R.

    o   load()ing an object saved in one of the formats used prior to
    R 1.4.0 is deprecated.  Such objects should be re-saved in the
    current format.

    o   save(version = 1) is now deprecated.


C-LEVEL FACILITIES

    o   The convenience function ScalarLogical now coerces all
    non-zero non-NA values to TRUE.

    o   The vector accessor functions such as INTEGER, REAL and
    SET_VECTOR_ELT now check that they are called on the
    correct SEXPTYPE (or at least on a compatible one).  See
    `Writing R Extensions' for the details and for a stricter
    test regime.

    o   It is no longer possible to pass list variables to
    .C(DUP = FALSE): it would have given rise to obscure garbage
    collection errors.

    o   allocString is now a macro, so packages using it will need to
    be reinstalled.

    o   R_ParseVector was returning with object(s) protected in the
    parser if the status was PARSE_INCOMPLETE or PARSE_ERROR.

    o   There is a new function Rf_endEmbeddedR to properly terminate
    a session started by Rf_initEmbeddedR, and both are now
    available on Windows as well as on Unix-alikes.  These and
    related functions are declared in a new header <Rembedded.h>.

    If R_TempDir is set when embedded R is initialized it is
    assumed to point to a valid session temporary directory: see
    `Writing R Extensions'.

    o   There is a new interface allowing one package to make C routines
    available to C code in other packages.  The interface consists
    of the routines R_RegisterCCallable and R_GetCCallable.  These
    functions are declared in <R_ext/Rdynload.h>. This interface
    is experimental and subject to change.

    In addition, a package can arrange to make use of header
    files in another (already installed) package via the
    'LinkingTo' field in the DESCRIPTION file: see 'Writing R
    Extensions'.


UTILITIES

    o   R CMD SHLIB now handles (as linker commands) -L*, -l* and *.a.

    o   R CMD check now:

    - warns if there are non-ASCII characters in the R code (as
      these will likely be syntax errors in some locale).

    - tests Rd cross-references by default, and tests for
      (syntactically) valid CITATION metadata.

    - tests that the package can be loaded, and that the package
      and name space (if there is one) can each be loaded in
      startup code (before the standard packages are loaded).

    - tests for empty 'exec' or 'inst' directories.

    - checks if $(FLIBS) is used when $(BLAS_LIBS) is.

    - checks that all packages (except non-S4-using standard
      packages) used in ::, :::, library() and require() calls are
      declared in the DESCRIPTION file, and 'methods' is declared if
      S4 classes or methods are set.

    - throws an error if the standard packages 'methods' and
      'stats4' are imported from in the NAMESPACE file and not
      declared in the DESCRIPTION file.

    o   The test script produced by massage-Examples.pl no longer
    creates objects in the base environment.

    o   New utilties R CMD Stangle and R CMD Sweave for extracting S/R
    code from and processing Sweave documentation, respectively.

    o   The DESCRIPTION file of packages may contain an 'Enhances:'
    field.

    o   An R CMD javareconf script has been added to allow Java
    configuration to be updated even after R has been installed.


INSTALLATION

    o   The C function realpath (used by normalizePath()) is hidden on
    some systems and we try harder to find it.

    o   There is a new option --enable-BLAS-shlib, which compiles the
    BLAS into a dynamic library -lRblas and links against that.
    For the pros and cons see the R-admin manual.

    The defaults are now --without-blas (so you have explicitly to
    ask for an external BLAS), and --enable-BLAS-shlib unless
    a usable external BLAS is found or on AIX or on MacOS X 10.2
    and earlier.

    o   MacOS X did not like having LSAME in both BLAS and LAPACK
    libraries, so it is no longer part of the R-internal LAPACK.
    We now require an external BLAS to provide LSAME: it seems
    that nowadays all do.

    o   The configure test for 'whether mixed C/Fortran code can be
    run' has been improved as on one system that test passed but
    the Fortran run-time library was broken.

    o   A precious configure variable DEFS can be set to pass defines
    (e.g. -DUSE_TYPE_CHECKING_STRICT) to C code when compiling R.

    o   There is now a test for visible __libc_stack_end on Linux
    systems (since it is not visible on some recent glibc's
    built from the sources).

    o   MacOS X 10.4 and higher now use two-level namespaces, single
    module in a shared library and allow undefined symbols to be
    resolved at run-time. This implies that common symbols are now
    allowed in package libraries. --enable-BLAS-shlib is supported
    for internal BLAS, external BLAS framework and external static
    BLAS.  An external dynamic library BLAS is NOT supported. (But
    it can be easily used by replacing internal BLAS library file
    later.)  MacOS X < 10.4 does not support --enable-BLAS-shlib.

    o   Dynamic libraries and modules use a flat namespace on MacOS X
    10.4 and higher if either Xcode tools don't support dynamic
    lookup (Xcode < 2.3) or the FORCE_FLAT_NAMESPACE environment
    variable is set.  (The latter was introduced temporarily for
    testing purposes and may go away anytime.)

    o   configure now defaults to 'run-time linking' on AIX
    (and AIX < 4.2 is no longer allowed), using -bexpall rather
    than export/import files.  If this works, it allows R to be
    built in the same way as other Unix-alikes, including with R
    as a shared library and with a shared BLAS.

    o   The "mac.binary" package type now defaults to universal
    binary.  If a repository supports architecture-specific Mac
    binaries, they can be requested by using "mac.binary.xxx" in
    contrib.url(), where xxx is the desired architecture.

    o   There is a new configure option --enable-memory-profiling to
        enable memory profiling with tracemem, Rprof, Rprofmem.


BUG FIXES

    o   The name of a Fortran symbol reported to be missing by
    .Fortran() is now the actual name.  (What was reported to be
    an 'entry point' was missing the common leading underscore.)

    o   print() on a MBCS character string now works properly a
    character at a time rather than a byte at time.  (This does
    not affect MBCSs like UTF-8 and the Windows DBCSes which have
    non-ASCII lead bytes and always worked correctly.)

    o   glm() now recalculates the null deviance whenever there is an
    offset (even if it is exactly zero to avoid a discontinuity in
    that case, since the calculations with and without offset are
    done by different algorithms).

    o   Amongst families, quasi() accepted an expression for link and
    no other did.  Now all accept an expression which evaluates to
    a one-element character vector (although e.g. 'logit' is taken
    as a name and not an expression).

    o   trace() now accepts arguments where= and signature= for the
    old-style trace (no tracer or exit, edit==FALSE) and just
    prints a message on entry.  Also the undocumented feature of
    where=function now works for generic functions as well.

    o   callNextMethod() failed for recursive use when the methods had
    nonstandard argument lists.  Now enforces the semantic rule
    that the inheritance is fixed when the method containing the
    callNextMethod() is installed.  See Details in the documentation.

    o   UseMethod() looked for the defining environment of 'generic' as
    if it were the current function, although some functions are
    generic for methods of a different generic.

    Lookup for S3 methods is confined to functions: previously a
    non-function 'fun.class' could have masked a function of the
    same name.

    o   Line types (lty) specified as hex strings were documented not to
    allow zero, but some devices accepted zero and handled it in a
    device-dependent way.  Now it is an error on all devices.
    (PR#8914)

    o   Subassignment for a time series can no longer extend the series:
    it used to attempt to but failed to adjust the tsp attributes.
    Now window() must be used.

    o   Function AIC() in package 'stats4' was not dispatching correctly
    on S4 classes via logLik() because of name space issues.

    o   Subsetting LANGSXPs could break the call-by-value illusion.
    (PR#7924) (patch from Kevin Hendricks).

    o   parse() with n > 1 gave a syntax error if fewer than n statements
    were available.

    o   parse() with n > 1 gave strange results on some syntax errors.
    (PR#8815)

    o   lag.plot() now respects graphical parameters for the axes.

    o   Using a wrong link in family() now gives more consistent error
    messages.

    o   sort.list(method="radix") works on factors again.

    o   object.size() is more accurate for vector objects (it takes into
    account the smaller header and also the fixed sizes used in
    the node classes for small vector objects).

    o   addmargins(T, ...) now returns a "table" when 'T' is a "table", as
    its help page has always suggested.

    o   remove() now explicitly precludes removing variables from
    baseenv() and throws an error (this was previously ignored).

    o   Saving the workspace at the end of a session now works as has
    long been intended, that is it is saved only if something has
    been added/deleted/changed during the current session.

    o   The search for bindings in <<-, ->> and assign(inherits=TRUE)
    was omitting the base package, although this was not
    documented.  Now the base package is included (but most
    bindings there are locked).

    o   dweibull(0, shape) was NaN not Inf for shape < 1.  Also, the
    help for dgamma and dweibull gave support as x > 0, but
    returned non-zero values for x = 0.  (PR#9080)

    o   Subsetting arrays no longer preserves attributes (it was
    removed for matrices in 1998).

    o   The "factor" method of as.character() no longer maps level
    "NA" to "<NA>" (a legacy of before there were NA character
    strings).

    o   terms(keep.order=TRUE) was not returning a valid "order"
    attribute.

    o   The DLL registration code was not freeing .External symbols.

    o   The internet download routines expected URLs of less than 4096
    bytes, but did not check.  Now this is checked, and http:// URLs
    are allowed to be up to 40960 bytes.

    o   parse(n=-1) threw a stack-imbalance error, and parse(n=3) did
    not cope correctly with EOF during input.

    o   Zero-column data frames had no names (rather than character(0)).

    o   by() and acf() could get confused when they used very long
    expressions as names.

    o   residuals(<glm object>, type="working") was NA for cases with
    zero weight (whereas they are well-defined even though the
    case was not used during the fitting) and the actual value is
    now returned.  This allows residuals to be computed from fits
    with 'y = FALSE'.

    The residuals in a fitted "glm" object are computed more
    accurately: the previous formula was subject to cancellation.

    o   loess() now checks the validity of its 'control' argument.

    o   rownames(<0-row matrix>, do.NULL=FALSE) was wrong.  (PR#9136)

    o   apply() now works as documented when applied over 2 or more
    margins with one of zero extent.  (It used to drop dimensions.)

    o   head() and tail() now also work row-wise for "table" and "ftable"
    objects.

    o   NextMethod() could throw an error/crash if called from a method
    that was called directly rather than from a generic (so
    .Method was unset).

    o   order(x, na.last = NA) failed for a zero-length x.

    o   grep(pat, x, value = TRUE, perl = L) preserved names for
    L == TRUE && !is.na(pat) but not otherwise.  Now it always does.

    o   [rc]bind() now find registered methods and not just visible ones.

    o   Printing a factor no longer ignores attributes such as names and
    dim/dimnames.

    o   Command-line arguments after --encoding were ignored.

    o   The check for impossible confidence levels was off by one in
    wilcox.test (PR#8557)

    o   [[ on an environment could create aliases. (PR#8457)

    o   pt() with a very small (or zero) non-centrality parameter could
    give an unduly stringent warning about 'full precision was not
    achieved'.  (PR#9171)

    o   writeChar() could segfault if 'nchars' was given silly values.

    o   qt() and rt() did not work for vector 'ncp', and qt() did not
    work for negative 'ncp'.

    o   ns() failed to work correctly when 'x' was of length one.

    o   identical() ignored tags on pairlists (including names of
    attributes) and required an identical ordering for
    attribute values in their pairlists.  Now names are compared
    on pairlists, and attribute sets are treated as unordered.

    o   If they were unused arguments supplied to a closure, only
    the first non-empty one was reported, despite the message.

    Unmatched empty arguments (such as f(1,,) for a function of
    one argument) were ignored.  They are now an error.

    o   Calling a builtin with empty arguments used to silently remove
    them (and this was undocumented).  Now this is an error unless
    builtin is c() or list() or there are only trailing empty
    arguments, when it is a warning (for the time being: this will
    be made an error in R 2.5.0).

    o   install.packages() ignored 'configure.args' if the vector was
    unnamed.

    o   biplot() now works if there are missing values in the data.

    o   biplot() now passes par() values to all four axes (not
    just those on sides 1 and 2).

    o   [.acf now handles an empty first index.

    o   Deparsing uses backticks more consistently to quote
    non-syntactic names.

    o   Assigning to the symbol in a for() loop with a
    list/expression/pairlist index could alter the index.  Now the
    loop variable is explicitly read-only.  (PR#9216)

    o   Using old.packages() (and hence update.packages()) on an empty
    (or non-existent) library failed with an obscure message.

    o   plot.xy() could segfault if supplied with an invalid 'col'
    argument.  (PR#9221)

    o   menu() with graphics=TRUE attempted to use Tcl/Tk on unix even
    if DISPLAY was not set (in which case Tk is not available and
    so the attempt is bound to fail).

    o   The print() method for 'dist' objects prints a matrix even for
    n = 2.

    o   The cumxxx functions were missing some PROTECTs and so could
    segfault on long vectors (especially with names or where
    coercion to numeric occurred).

    o   The X11() device no longer produces (apparently spurious)
    'BadWindow (invalid Window parameter)' warnings when run from
    Rcmdr.

    o   legend() assumed that widths and heights of strings were positive,
    which they need not be in user coordinates with reversed axes.
    (In part, PR#9236)

    o   The plot() methods for "profile.nls" objects could get
    confused if 'which' had been used in the profile() call. (PR#9231)

    o   boxplot() did not passed named arguments (except graphics
    parameters) to bxp() as documented.  (PR#9183)

    o   Only genuinely empty statements act as 'return' in the
    browser, not say those starting with a comment char.  (PR#9063)

   o    summary.mlm() incorrectly used accessor functions to fake an
    "lm" object.  (PR#9191)

   o    prettyNum() was not preserving attributes, despite being
    explicitly documented to. (PR#8695)

   o    It was previously undocumented what happened if a graphical
    parameter was passed in both '...' and 'pars' to boxplot()
    and bxp(), and they behaved differently.  Now those passed in
    '...' have precedence in both cases.

    o   A failed subassignment could leave behind an object '*tmp*'.
    The fix also sometimes gives better error messages.

    o   Using SIGUSR1 on Unix now always terminates a session, and no
    longer is caught by browser contexts and restarts (such as try()).

    o   In the 'graphics' package, in-line 'font=5' was being ignored
        (report by Tom Cook).

    o   nls() looked for non-parameter arguments in a function call in
    the wrong scope (from the body of nls).

    o   Printing of complex numbers could misbehave when one of the
    parts was large (so scientific notation was used) and the
    other was so much smaller that it had no significant digits
    and should have been printed as zero (e.g. 1e80+3e44i).

    o   Using install.packages with type="mac.binary" and target path
    starting with ~ failed with a cryptic message while unpacking.

    o   getwd() now works correctly when the working directory is
    unavailable (e.g. unreadable).

    o   The alternative hypothesis in wilcox.test() was labelled by an
    unexplained quantity 'mu' which is now spelled out.

    The alternative hypothesis in ks.test() is clearer both in the
    documentation and in the result. (PR#5360)



    **************************************************
    *                        *
    *          2.3 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 2.3.1 patched


BUG FIXES

    o   help.search() incorrectly documented where the data base was
    saved and what was in it.

    o   str(<data.frame>, strict.width=.) produced an extraneous first line.

    o   Improve bessel[IJKY]() out-of-range checks; for instance,
    besselI(0, nu) no longer warns.

    o   cbind() on some non-vector objects (e.g. names) segfaulted.

    o   xy.coords(numeric(0)) gave a misleading error message.

    o   INSTALL has been changed to avoid a few problems with the
    base (not XPG4) Solaris sed.

    o   There were problems in src/main/printutils.c on some platforms
    that did not have va_copy.

    o   [pq]unif allowed infinite ranges but handled them
    inconsistently.  (PR#8958)

    [pq]unif claimed to allow min == max but did not do so.

    o   Subassignment within raw arrays of 3 or more dimensions was not
    implemented.

    o   environment<-() did not even check if duplication was required.

    o   rt() was wrong for non-central t-distributions. qt() was wrong
    for non-central distributions far enough in the left tail
    (reported by Long Qu).  (PR#9050)

    o   R CMD Rdconv --type=Ssgm no longer drops a \keyword{} if it is the
    only one (one letter typo patch by Bill Dunlap).  (PR#9051)

    o   poly(x, y, .., raw = TRUE) now does follow 'raw'.

    o   plot(0/ -1:1, type = "s") now works.  (PR#9046)

    o   write.table(row.names=FALSE) does now quote column names. (PR#9044)

    o   rf() was wrong for non-central F (PR#9055). qf() was wrong for
    non-central F with denominator df > 1e8.

    o   The C function dummy_vfprint had a potential memory leak.  This
    showed up with some connections when used with output of more
    than 100,000 bytes in a single call.

    o   vignette() did not work if more than one vignette with the
    same name was installed  (PR#9069).

    o   mget() did not check that 'ifnotfound' was a list, and worked
    incorrectly if it was a list of length > 1.  Now 'ifnotfound'
    is (if possible) coerced to a list.

    o   The internal check for conformance of time series assumed that the
    'tsp' attribute was integer, whereas it was enforced to be
    numeric by the attribute-setting code.

    o   Sweave() did not treat \begin{document} in comments correctly
    (PR#9073).

    o   diff(x) gave wrong result when x was "POSIXlt" object.

    o   title() could crash R when passed a long non-character vector.
    (PR#9115)

    o   Functions rgb(), hsv() and hcl() lost names unless 'alpha' was
    specified.  (In part, PR#9118)

    o   The residuals() method for "glm" failed unless 'y = TRUE' for
    the fit: it now works for "working" and "partial" residuals
    and gives a useful error message for other types. (PR#9124)

    o   R CMD Rdconv did not close the .Rd file, which matters on Windows.
    (PR#9126)

    o   anova.mlm() could fail if applied to a single model.

    o   stopifnot(<FALSE-expr>) now gives a better message for a *long*
    expression.




        CHANGES IN R VERSION 2.3.1


NEW FEATURES

    o   In lm() and glm(), offsets are allowed to be length 1 (and
    if so are replicated to the number of cases).

    o   The \uxxxx notation for Unicode characters in input strings
    can now be used on any platform which supports MBCS, even if
    the current locale is not MBCS (provided that the Unicode
    character is valid in the current character set).

    o   The quasibinomial() family now allows the "cauchit" link. (PR#8851)

    o   edit.data.frame() no longer (silently) coerces character columns
    to factor.


C-LEVEL FACILITIES

    o   The variables controlling stack checking are made available via
    Rinterface.h to front-ends embedding R: see 'Writing R Extensions'

    o   R_SignalHandlers (defined in Rinterface.h) can be set to 0 to
    suppress the R signal handlers in front-ends embedding R.


INSTALLATION CHANGES

    o   There have been a number of changes to help installation on
    platforms that no one had beta-tested.

    - Changes related to older header files, e.g. on Redhat 8.0/9.

    - Problems with 'make install' on older (<3) versions of bash
      on Solaris and elsewhere.

    - AIX 5.2/gcc issues with needing -lm when making modules X11
      and vfonts.

    - Some versions of Solaris and AIX had an fcntl.h that
      redefined 'open' to be 'open64' and thereby broke
      compilation of src/main/connections.c and elsewhere.

    o   'make uninstall' works better on a build using a named
    subarchitecture.


BUG FIXES

    o   min(), max(), sum() and prod() gave nonsensical answers with
    an empty list or raw argument.

    o   sum() on a data frame did not allow multiple arguments.  (PR#8385)

    o   charmatch() and pmatch() did not specify they applied only to
    character vectors.  Now they do, and attempt to coerce 'x' and
    'target' to character before attempting matching.

    o   The Summary() methods for data.frame, Date, POSIXct, POSIXlt
    and difftime all required an argument which can match 'x',
    although the generics did not.

    o   regexpr() now accepts 0-length 'text' inputs.

    o   help.search() no longer errors out on a wrongly installed package
    (with no "hsearch.rds" file).

    o   The LaTeX version of the package reference manual was omitting
    some topics, and was not sorting the foo-package topic first.

    o   Serializing (e.g. via save()) is better protected against C
    stack overflow, which will now abort the conversion but no
    longer crashes the R process on some platforms.

    o   rbind()ing dataframes with a single row could lead to a
    corrupt data frame (a problem with the fix to PR#8506).

    o   plot(lm(y ~ 1)) now works also for 'which = 5'.

    o   dbeta(0, 1, a, 0) now correctly gives 'a' (limit) instead of 0,
    and dbeta(0, a, b, ncp) now returns Inf instead of NaN.

    o   demo(Hershey) was failing on the Cyrillic octal codes in locales
    (e.g. UTF-8) in which these are invalid.

    o   mean() on an integer (or logical) vector was treating NAs as
    actual values (unless na.rm = TRUE).

    o   mean() on a complex vector was calculated incorrectly in
    code to improve precision (PR#8842, John Peters).

    o   Graphical parameters bg, cex, col, lty, and lwd were being
    checked as being of length one even by functions such as
    title() that ignored them.  (Functions such as lines() and
    points() allow them to be of length > 1, so they might be
    passed through ... to other high-level graphical functions
    which then used to reject them.)

    o   str() now is fast again for large character vectors.

    o   edit() would default the environment of a function to .BaseEnv,
    instead of to .GlobalEnv.

    o   lm() and glm() coerce their 'weights' and 'offset' values to
    vector to avoid problems with specifying them as 1D or n x 1
    arrays.

    o   image() with one or both axes on log scales would give a
    spurious warning; contour() would give an error.

    o   legend() with log axes would place the title in the wrong place.

    o   edit.data.frame() was not returning factors edited with
    factor.mode="numeric" to factors.

    o   edit.matrix() tried to set rownames and colnames from the
    original matrix even if the sizes had been altered, and
    ignored changes made to the column names.  'edit.row.names' has
    a more sensible default (if the rownames are non-NULL).

    o   bindingIsLocked() was returning invalid values of a logical
    vector on some platforms.

    o   merge.data.frame() did not make the column names unique (by
    appending elements of 'suffixes') when performing a Cartesian
    product.  (PR#8676)

    o   rbind.data.frame() matches up the names of columns (which was
    undocumented), but failed to do so when checking if it was
    dealing with a factor column.  (PR#8868)

    If rbind() was used on data frames with duplicated names it
    produced a corrupt data frame.

    o   dt(x, df, ncp= not.0) no longer gives erratic values for
    |x| < ~1e-12.  (PR#8874)

    o   \code{\linkS4class{.}} now works.

    o   ccf() aligns time series by ts.intersect() rather than
    ts.union() and so is less likely to need a non-default
    na.action.  (PR#8893)

    o   optim(method="CG") could return a value that did not
    correspond to $par for very badly behaved functions on which
    the second phase of the line search failed.  (PR#8786)

    o   print.ts() could fail on a corrupt time series: it now warns and
    does the best it can.




        CHANGES IN R VERSION 2.3.0


USER-VISIBLE CHANGES

    o   In the grid package there are new 'arrow' arguments to
    grid.line.to(), grid.lines(), and grid.segments()
    (grid.arrows() has been deprecated).

    The new 'arrow' arguments have been added BEFORE
    the 'name', 'gp' and 'vp' arguments so existing code that
    specifies any of these arguments *by position* (not by name)
    will fail.

    o   all.equal() is more stringent, see the PR#8191 bug fix below.

    o   The data frame argument to transform() is no longer called 'x',
    but '_data'.  Since this is an invalid name, it is less likely
    to clash with names given to transformed variables. (People
    were getting into trouble with transform(data, x=y+z).)


NEW FEATURES

    o   arima.sim() has a new argument 'start.innov' for compatibility
    with S-PLUS.  (If not supplied, the output is unchanged from
    previous versions in R.)

    o   arrows() has been changed to be more similar to segments():
    for example col=NA omits the arrow rather than as previously
    (undocumented) using par("col").

    o   as.list() now accepts symbols (as given by as.symbol() aka
    as.name()).

    o   atan2() now allows one complex and one numeric argument.

    o   The 'masked' warnings given by attach() and library() now only
    warn for functions masking functions or non-functions masking
    non-functions.

    o   New function Axis(), a generic version of axis(), with Date and
    POSIX[cl]t methods.  This is used by most of the standard
    plotting functions (boxplot, contour, coplot, filled.contour,
    pairs, plot.default, rug, stripchart) which will thus label x
    or y axes appropriately.

    o   pbeta() now uses TOMS708 in all cases and so is more accurate
    in some (e.g. when lower.tail = FALSE and when one of the
    shape parameters is very small).

    o   [qr]beta(), [qr]f() and [qr]t() now have a non-centrality parameter.

    o   [rc]bind and some more cases of subassignment are implemented
    for raw matrices.  (PR8529 and 8530)

    o   The number of lines of deparsed calls printed by browser() and
    traceback() can be limited by the option "deparse.max.lines".
    (Wish of PR#8638.)

    o   New canCoerce() utility function in "methods" package.

    o   [pq]chisq() are considerably more accurate for moderate (up to
    80) values of ncp, and lower.tail = FALSE is fully supported
    in that region.  (They are somewhat slower than before.)

    o   chol(pivot = TRUE) now gives a warning if used on a (numerically)
    non-positive-definite matrix.

    o   chooseCRANmirror() consults the CRAN master (if accessible) to
    find an up-to-date list of mirrors.

    o   cov.wt() is more efficient for 'cor = TRUE' and has a new 'method'
    argument which allows 'Maximum Likelihood'.

    o   do.call() gains an 'envir' argument.

    o   eigen() applied to an asymmetric real matrix now uses a
    tolerance to decide if the result is complex (rather than
    expecting the imaginary parts of the eigenvalues to be exactly
    zero).

    o   New function embedFonts() for embedding fonts in PDF or
    PostScript graphics files.

    o   fisher.test() now uses p-values computed via hypergeometric
    distributions for all 2 by 2 tables.  This might be slightly
    slower for a few cases, but works much better for tables with
    some large counts.

    There is a new option to simulate the p-value for larger than
    2 x 2 tables.

    o   for() now supports raw vectors as the set of indices.

    o   getNativeSymbolInfo() is vectorized for the 'name' argument.  It
    returns a named list of NativeSymbolInfo objects, but is
    backward compatible by default when called with a character
    vector of length 1, returning the NativeSymbolInfo object.

    o   help.search() no longer attempts to handle packages installed
    prior to R 2.0.0, and reports the current path to the
    package (rather than where it was originally installed: this
    information is not shown by the print() method).

    o   Added "hexmode" to parallel "octmode".

    o   install.packages() now does tilde expansion on file paths
    supplied as 'pkgs'.

    o   install.packages() has additional arguments 'configure.args' and
    'clean' which allow the caller to provide additional arguments
    to the underlying R CMD INSTALL shell command when installing
    source packages on a Unix-alike.

    o   is.loaded() has a new argument 'type' to confine the search to
    symbols for .C, .Fortran, .Call or .External: by default it
    looks for a symbol which will match any of them.  It is now
    internal and not primitive, so argument matching works in the
    usual way.

    o   The symmetry test for matrices used in eigen() has been ``exported''
    as the 'matrix' method of a new S3-generic 'isSymmetric().

    o   .leap.seconds and the internal adjustment code now know about
    the 23rd leap second on 2005-12-31: the internal code uses a
    run-time test to see if the OS does.

    o   The 'col' argument of legend() now defaults to par("col")
    (which defaults to "black", the previous default), so that
    the lines/symbols are shown in the legend in the colour that
    is used on the plot.

    o   log2() and log10() call C functions of the same name if
    available, and will then be more likely to be precise to
    machine accuracy.

    o   new.packages() gains a ... argument to pass e.g. 'destdir' to
    install.packages().  (Wish of PR#8239.)

    o   nls() now supports 'weights'.

    o   The vector passed as the first argument of the 'fn' and 'gr'
    arguments of optim() has the names (if any) given to argument
    'par'.

    o   options(expressions) is temporarily increased by 500 during
    error-handling.   This enables e.g. traceback() to work when
    the error is reaching the limit on the nesting of expressions.

    o   page() accepts general R objects, not just names (and
    previously undocumented) character strings.  This allows the
    object to be specified as a call, for example.  More options
    are allowed in its '...' argument.

    o   pairs() allows a wider class of inputs, including data frames
    with date and date-time columns.

    o   par() and the in-line use of graphical parameters produce more
    informative error messages, distinguishing between
    non-existent pars and inappropriate use of valid pars.

    Graphical parameters 'family', 'lend', 'ljoin' and 'lmitre'
    can now be set in-line.

    There is no longer a warning if non-settable pars are used
    in-line, but there is an appropriate warning if unknown pars
    are passed.

    The length limit for the 'family' parameter has been increased
    to 200 bytes, to allow for the names of some CID-keyed fonts
    in multi-byte locales.

    o   The pdf() device now allows 'family' to be specified in the same
    generality as postscript().

    o   The pdf() device writes /FontDescriptor entries for all fonts
    except the base 14, and does not write font entries for unused
    fonts.

    o   Plotmath allows 'vartheta', 'varphi' and 'varsigma' (or 'stigma')
    as synonyms for 'theta1', 'phi1' and 'sigma1', and the help
    page has a note for TeX users.

    o   plot.xy() now takes its default arguments from the
    corresponding par() settings, so points(type="l") and
    lines(type="p") behave in the same way (and more obviously,
    also for type="b").

    o   poly() has a new argument 'raw', mainly for pedagogical purposes.

    o   The class "POSIXlt" now supports fractional seconds (as
    "POSIXct" has always done).  The printing of fractional seconds
    is controlled by the new option "digits.secs", and by default is off.

    o   postscript() supports family = "ComputerModernItalic" for
    Computer Modern with italic (rather than slanted) faces.

    o   The postscript()/pdf() font metrics for the 14 standard fonts
    (only, not the rest of the common 35) have been updated to
    versions from late 1999 which cover more glyphs.  There are
    also a few differences in the metrics and hence the output
    might be slightly different in some cases.

    o   The way families can be specified for postscript() and pdf()
    has been expanded to include CID-keyed fonts, with new
    functions Type1Font() and CIDFont() to set up such fonts
    families.

    o   prettyNum() has new arguments 'preserve.width' and 'zero.print'.
    When the former is not "none", as in calls from format() and
    formatC(), the resulting strings are kept at the desired width
    when possible even after adding of 'big.mark' or 'small.mark'.

    o   proc.time() and system.time() now record times to 1ms accuracy
    where available (most Unix-like systems).

    o   The initialization methods for the quasi() family have been
    changed to depend on the variance function, and in particular
    to work better for the "mu(1-mu)" variance function.  (PR#8486)

    o   read.table() gains a 'flush' argument passed to scan().

    o   require() now takes a 'lib.loc' argument.

    o   The second argument 'size' to sample() is required to have
    length 1, so that errors when supplying arguments are more
    easily detected.

    o   The default is now compress = !ascii in save() (but not save.image).

    o   scan() and write.table() now have some interruptibility, which
    may be useful when processing very large files.

    o   A new heuristic test, seemsS4Object() is supplied, along with a
    similar C-level test, R_seemsS4Object(object). The test
    detects probable S4 objects by their class's attribute. See
    the help page.

    o   S3 classes can now be made non-virtual S4 classes by supplying a
    prototype object in the arguments to setOldClass().

    o   splinefun() returns a function that now also has a 'deriv' argument
    and can provide up to the 3rd derivative of the interpolating
    spline, thanks to Berwin Turlach.

    o   stopifnot(A) now gives a better error message when A has NAs, and
    uses "not all TRUE" when A has length >= 2.

    o   str()'s default method has a new argument 'strict.width' which can
    be used to produce strict 'width' conforming output.  A new
    options(str = list(strict.width = *)) setting allows to control
    this for a whole session.

    o   summary.nls() has a new argument 'correlation' that defaults
    to FALSE (like summary.lm).

    o   Sys.sleep() has sub-millisecond resolution on Unix-alikes
    with gettimeofday().

    o   Sys.time() now has sub-millisecond accuracy on systems
    supporting the POSIX call gettimeofday, and clock-tick
    accuracy on Windows.

    o   The new function timestamp() adds a time stamp to the saved
    command history on consoles which support it.

    o   New function tcrossprod() for efficiently computing x %*% t(x)
    and x %*% t(y).

    o   The suffix used by tempfile() is now in hex on all platforms
    and guaranteed to be at least 6 hex digits (usually 8).

    o   trace() now works more consistently and more like its
    documentation, in particular the assertions about old tracing
    being removed for new.  For debugging purposes (of R) a
    mechanism for debugging the trace computations themselves was
    added.  See trace.R.

    o   The implementation of trace() has beem made more general by
    calling a function to do the trace interaction, and recover()
    now detects trace calls to trim the irrelevant code underneath.

    o   unserialize() can now also read a byte stream from a raw vector.

    o   The useDynLib() directive in the NAMESPACE file now accepts the
    names of the native/foreign symbols that are to be resolved in
    the DLL for use in .C/.Call/.Fortran/.External calls.  These
    can be used as regular R variables instead of the (routine
    name, PACKAGE) pairs currently recommended.  Alternative names
    can be given for the R variables mapping to these symbols.
    The native routine registration information can also be used
    directly via useDynLib(name, .registration = TRUE).  See the
    'Writing R Extensions' manual for more details.

    checkFF() (package 'tools') has been updated accordingly.

    o   validObject() has an option complete=TRUE that recursively
    checks the objects in the slots.  Not used when new(...)
    checks validity.

    o   New Vectorize() function, a wrapper for mapply().

    o   write.ftable() has gained an argument 'append = FALSE' (thanks to
    Stephen Weigand).

    o   On Unix-alikes, X11() now has arguments to request the initial
    position of the window, and 'gamma' defaults to the value of
    getOptions("gamma").  These changes are consistent with the
    windows() device.

    o   X11() and the Unix-alike data entry window can have properties
    (including geometry) set by X resources: see their help files.

    o   xy.coords() & xyz.coords() now have NULL defaults for their 'y' or
    'y' and 'z' arguments.  This is more consistent with their earlier
    documentation, and may be convenient for using them.


    o   Non-syntactic names of list elements are now printed quoted by
    backticks rather than double quotes.

    o   There is some basic checking for imminent C stack overflow (when
    the evaluation depth and the user interrupts are checked).
    On systems with suitable OS support (not Windows), segfaults
    from C stack overflow are caught and treated as an R error.

    New function Cstack_info() reports on stack size and usage.

    options(expressions) reverts to the default of 5000 now
    stack checking is in place.

    o   Package tcltk does not try to initialize Tk on Unix-alikes
    unless a DISPLAY variable is present.  This allows packages
    dependent on tcltk to be installed without access to an X server.

    o   The code used to guess timezone offsets where not supplied by
    the OS uses a different algorithm that is more likely to guess
    the summer-time transitions correctly.

    o   Package tools contains translation tables 'Adobe_glyphs' and
    'charset_to_Unicode'.

    o   Changed the environment tree to be rooted in an empty
    environment, available as emptyenv().  baseenv() has been
    modified to return an environment with emptyenv() as parent,
    rather than NULL.

    o   gettext has been updated to 0.14.5.

    o   PCRE has been updated to version 6.4.

    o   The method $.DLLInfo resolves the specified symbol in the DLL,
    returning a NativeSymbolInfo object.  Use [[ to access the actual
    values in the DLLInfo object.

    o   On systems with either vasprintf or both va_copy and a vsnprintf
    which reports the size of buffer required, connections such as
    gzfile() and bzfile() can now write arbitrarily long lines,
    not just 100000 chars.

    o   The R session temporary directory is now set in C code using
    the same algorithm whether or not the shell front-end is used
    and on all platforms.  This looks at environment variables
    TMPDIR, TMP and TEMP in turn, and checks if they point to a
    writable directory.

    o   Some of the classical tests put unnecessary restrictions on the
    LHS in the formula interface (e.g., t.test(x+y ~ g) was not
    allowed).

    o   On suitably equipped Unix-alike systems, segfaults, illegal
    operations and bus errors are caught and there is a simple
    error-handler which gives the user some choice as to what to
    do in interactive use. [Experimental.]

    On Windows access violations and illegal instructions are
    caught with a simple error handler.  [Experimental.]

    o   Tracebacks now include calls such as .C/.Fortran/.Call, which
    will help if errors occur in R code evaluated by compiled code
    and in tracebacks presented by the segfault etc handlers.

    o   Treatment of signature objects and method definition objects has
    been modified to give cleaner printing and more consistency in
    the treatment of signatures.  A sometimes useful utility,
    methodSignatureMatrix(), is now exported.

    o   Printing the results of codoc() from package tools now helpfully
        summarizes the found code/documentation mismatches.

    o   R refrains from printing a final EOL upon exiting the main loop
    if the quiet flag is on and if the save action is known (e.g. this
    is true for --slave).


DEPRECATED & DEFUNCT

    o   The deprecated and undocumented use of atan() with two arguments
    has been removed: instead use atan2().

    o   write.table0() is defunct in favour of write.table().

    o   format.char() is defunct in favour of format.default().

    o   Support for the long-deprecated (and no longer documented)
    arguments --min-vsize --min-nsize --max-vsize --max-nsize
    --vsize --nsize of R CMD BATCH has been removed.

    o   The 'debian' subdirectory has been removed from the sources.

    o   The 'vfont' argument of axis() and mtext() has been removed:
    use par(family=) instead.

    o   The unused graphical parameter "type" has been removed: it
    invited confusion with the 'type' argument to default methods
    of plot(), points() and lines().

    o   nlsMethod() and profiler() are no longer exported from the stats
    name space (and nlsMethod.plinear() is no longer registered as
    a method, as nlsMethod() was not generic).

    o   The re-named tcltk functions tkcmd, tkfile.tail, tkfile.dir,
    tkopen, tkclose, tkputs, tkread are now formally deprecated.

    o   Argument 'col' of bxp() is now formally deprecated.

    o   Use of NULL as an environment is deprecated and gives a warning.

    o   postscriptFont() is deprecated in favour of Type1Font() (which
    is just a change of name).

    o   La.chol() and La.chol2inv() are deprecated (they have since R
    1.7.0 been the same as the default options of chol() and chol2inv).

    o   La.svd(method = "dgesvd") is deprecated.

    o   The use of install.R and R_PROFILE.R files in packages is
    deprecated: use the DESCRIPTION file instead to arrange to
    save an image or to load dependent packages.

    The following command-line options to INSTALL are deprecated
    (use the fields in the DESCRIPTION file instead):
    -s --save --no-save --lazy --no-lazy --lazy-data --no-lazy-data

    o   Graphical parameter 'tmag' (which is long unused) is deprecated.


INTERNATIONALIZATION

    A set of patches supplied by Ei-ji Nakama has been incorporated.

    o   New postscript encodings for CP1253, CP1257 and Greek (ISO 8859-7).

    o   Support for East Asian CID-keyed fonts in pdf() and postscript().
    Although these usually contain Latin characters no accurate
    AFMs are available and so CID-keyed fonts are intended only
    for use with CJK characters.

    o   Wide-character width functions wc[s]width are provided that
    overcome problems found with OS-supplied ones (and those
    previously used by R on Windows).  This means that
    double-width CJK characters are now supported on all
    platforms.  It seems that the width of some characters (and
    not just CJK characters) depends on which CJK locale's fonts
    are in use and also on the OS.

    Revised wide-character classification functions are
    provided for use on Windows, AIX and MacOS X to replace
    deficient OS-supplied ones.

   o    There is support for MBCS charsets in the pictex() graphics
    device, and rotated (by 90 degrees) text may work better.

   o    The \u (and \U except on Windows) notation for characters which
    is supported by the parser in all MBCS charsets is now always
    interpreted as a Unicode point, even on platforms which do not
    encode wchar_t in Unicode.  These are now a syntax error in
    single-byte locales.

    o   The default encoding for postscript() and pdf() is chosen to
    be suitable for the current locale, if that is a single-byte
    locale which is supported.  This covers European (including
    Greek) and Cyrillic languages.

    In UTF-8 locales, a suitable single-byte encoding is chosen
    for postscript() and pdf(), and text translated to it.

    o   xfig() gains an 'encoding' argument.

    o   There are some message translations into Spanish.


INSTALLATION CHANGES

    o   The encoding files for pdf()/postscript() have been moved to
    directory 'enc' in package 'grDevices'.

    o   Support for MBCS is only enabled if iconv is found and it
    supports enough conversions.  (libiconv does.)

    o   In an MBCS locale, make check now translates the graphics
    examples from Latin-1.  This ensures that they will work
    correctly in UTF-8: it is possible that in other MBCS locales
    they will now fail (rather than work completely incorrectly).

    o   There is a new test, 'test-Docs', which as part of 'make
    check-devel' tests the code in the documentation.  Currently
    it runs doc/manual/R-{exts,intro}.R and the compiled code in
    R-exts.c.

    o   The workaround to allow an external LAPACK-containing BLAS
    such as libsunperf to be used with the internal LAPACK has
    been removed.  If you have such a library you may now need
    to use --with-lapack.  It is no longer possible to use some
    older versions of libsunperf, e.g. Forte 7 on 64-bit builds.

    o   A substitute for mkdtemp is provided, so it is now always
    used for R_TempDir.

    o   Most of the functions checked for by 'configure' also have
    declarations checked for in the appropriate header.

    o   The top-level documentation files AUTHORS COPYING.LIB COPYRIGHTS
    FAQ RESOURCES THANKS have been moved to doc, and COPYING and NEWS
    are installed there.  The file Y2K has been removed from the
    distribution.

    o   The extension .lo is no longer used in building R (only in
    the optional build of libRmath.so): this allows a considerable
    simplification of the Makefiles.

    o   Direct support for f2c has been removed: it can still be used
    via a script which makes it look like a Fortran compiler.
    (src/scripts/f77_f2c is an example of such a script.)

    o   There is a new flag SAFE_FFLAGS which is used for the
    compilation of dlamc.f.  It is set by configure for known
    problem cases (recent g77 and gfortran), but can be overridden
    by the user.

    o   The standard autoconf macros for large-file support are now used,
    and these are enabled unless --disable-largefile is specified.
    This replaces --enable-linux-lfs (and is now selected by default).

    o   Visibility attributes are used where supported (gcc4/gfortran on
    some platforms, also gcc3/g77 on FC3 and partially elsewhere).
    The main benefit should be faster loading (and perhaps better
    optimized code) in some of the dynamic shared objects (e.g.
    libR.so and stats.so).

    o   The *PICFLAGS are taken to be -fpic rather than -fPIC where
    possible.   This will make no difference on most platforms:
    -fPIC is needed on Sparc (and still used there), but -fpic
    should give slightly better performance on PowerPC (although
    -fPIC is used on PPC64 as it is needed to build libR.so there).

    o   More use is made of inlining for small utility functions such
    as isReal.  Because this can only be done portably with C99
    constructs (and we know of no actual implementation), this is
    only done for the GNU C compiler.

    o   There is an experimental feature to allow shared installations
    of sub-architectures.  See the R-admin manual.

    o   All platforms now use R's internal implementation of strptime,
    which allows fractional seconds.  (The major platforms were
    already using it.)

    o   The dlcompat work-around for old Mac OS X systems (<= 10.2)
    has been removed.  External dlcompat must be installed if needed.


UTILITIES

    o   R CMD check now uses an install log by default.

    o   R CMD check works for packages whose package name is different
    from the directory name in which it is located.

    o   R CMD INSTALL now uses more randomness in the temporary
    directory name even on systems without mktemp -d.

    o   R CMD f77 has been removed now f2c is no longer supported.

    o   The version string shown in the startup message and by
    "R --version", and that stored in variable R.version.string
    are now in exactly the same format.

    o   The base name of a help file needs to be valid as part of a file://
    URL, so R CMD check now checks the names are ASCII and do not
    contain % .

    o   R CMD check now warns about unknown sections in Rd files, and
    invalid names for help, demo and R files, as well as unlikely
    file names in the 'src' directory.  The latter is controlled
    by option --check-subdirs and by default is done if checking a
    tarball without a configure script.

    R CMD build excludes invalid files in the 'man', 'R' and 'demo'
    subdirectories.

    o   \usepackage[noae]{Sweave} in the header of an Sweave file
    supresses auto-usage of the ae package ("almost European"
    fonts) and T1 input encoding.


DOCUMENTATION

    o   Rd format now allows \var{} markup inside \code{} and \examples{}.

    o   Markup such as --, ---, < and > is handled better when
    converting .Rd files to [C]HTML.

    o   There is new markup \link[=dest]{name} to generate a link to topic
    'dest' which is shown as 'name', and \linkS4class{abc} which
    expands to \link[=abc-class]{abc}, for cross-referencing the
    recommended form of documentation for S4 classes.


PACKAGE INSTALLATION

    o   There is now some support for Fortran 90/95 code in packages:
    see `Writing R Extensions'.

    o   Installation of man sources and demos is now done by R code.
    The restrictions on the names of help files, R files and
    of demos are now enforced (see `Writing R Extensions').

    o   Packages which contain compiled code can now have more than
    one dot in their name even on Windows.

    o   The Meta/hsearch.rds database saved now contains LibPath="".
    This information is now always recreated when help.search() is
    run, but the field is retained for back-compatibility.

    o   update.packages() now has a '...' argument to be passed to
    install.packages(), including the formerly separate arguments
    'destdir' and 'installWithVers'.

    o   Make macros AR and RANLIB are now declared in etc/Makeconf for
    use by packages which wish to make static libraries.


C-LEVEL FACILITIES

    o   qgamma and rgamma in Rmath.h now check for non-positive arguments.

    o   The BLAS which ships with R now contains the complete set of
    double-complex BLAS routines, rather than just those used in R.

    <R_ext/BLAS.h> has been corrected to add the missing
    double-precision BLAS functions drotmg and drotm, and to
    exclude lsame (which is a Lapack auxiliary function and is now
    declared in <R_ext/Lapack.h>).  It also includes the double
    complex routines added for this release of R provided Fortran
    doublecomplex is usable on the platform.

    o   <R_ext/BLAS.h> and <R_ext/Lapack.h> now declare all the entry
    points as 'extern'.

    o   The flag SAFE_FFLAGS is made available to packages via
    etc/Makeconf and R CMD config.  It can be used where
    optimization needs to be defeated, e.g. in LAPACK setup.

    o   getNativeSymbolInfo has a withRegistrationInfo argument which
    causes the address field to be a reference to the registration
    information if it is available for that symbol.  If the
    registration information is not available, the address is a
    reference to the native symbol. The default is FALSE which is
    backward compatible, returning just the address of the symbol
    and ignoring registration information.

    o   errorcall and warningcall are now declared in <Rinternals.h>
    (they might be needed in front-ends).

    o   R_FlushConsole and R_ProcessEvents are now declared in <R.h>.

    o   The R_Sock* functions supporting socket connections are no
    longer declared in R-ftp-http.h as they are not loaded into R
    itself, and are now hidden in the module's DLL on suitable systems.


BUG FIXES

    o   Quoted arguments to the R script after --args are now
    passed quoted to the R executable and so will be shown as
    expected by commandArgs().  (They were previously split at
    whitespace even inside quotes on Unix-alikes but not on
    Windows.)

    o   axis() now supports pars 'xaxp'/'yaxp' as inline arguments.

    o   sort() now does not return inappropriate attributes such as
    "dim" and "tsp": it only returns names.

    sort(x, partial=) no longer returns unsorted names, and drops
    names (since it is supplied for efficiency).

    o   Use of non-central F in pf() gives accurate values for larger ncp.

    o   R CMD build --binary does a better job of cleaning up after
    failure to re-make vignettes.

    o   reg-test-1.R tested system(intern=TRUE) which depends on popen
    and so is not supported on all platforms.

    o   Changed apparent mis-spelling of "Gibraltar" in dataset 'eurodist'.

    o   sysconf() is now used to find the number of clock ticks/second:
    under some circumstances glibc reported CLK_TCK = 60 when the
    true value was 100.

    o   identical() was not allowing for embedded nuls in character
    strings. (NB: the comparison operators including == do not,
    and never will.)

    o   The profile() and profiler() methods for "nls" objects now
    support algorithm = "plinear" and algorithm = "port".

    o   The signal handlers for signals USR1 and USR2 where not
    restored if the signal arrived when interrupts were suspended.

    o   Certain combinations of S4 inheritance could cause inherited
    methods to override some directly specified methods.

    o   Some cases of named signatures in calls to setMethod() caused errors.

    o   all.equal() is now more consistent and "picky" about mismatching
    attributes, in particular names(); this is a part of the
    propositions by Andy Piskorkski (PR#8191).

    o   load() when applied to a connection leaves it open/not as it
    found it, and checks explicitly for having a binary readable
    connection.

    o   The p-values given by stat.anova() (called from several
    anova() methods) are now NA (rather than spurious) if
    non-nested models give rise to changes in deviance with a
    different sign from changes in degrees of freedom.

    o   Built-ins were reported as the relevant call in C-level error()s
    iff R profiling was in progress.  Now they are never reported.

    o   Too-long signatures (with no names) were not being caught
    in setMethod().

    o   Slot names in prototype() are being more thoroughly checked.

    o   signif() is more likely to follow the 'round to even' rule for
    exactly representable numbers, e.g. signif(0.25, 1).  (Related
    to PR#8452.)

    o   nls() now works correctly with some low-dimensional fits,
    e.g. with one or zero non-linear parameters.

    o   glm() could give an inappropriate error message if all
    possible coefficients were invalid (e.g. a log-linear binomial
    model with no intercept and a not all positive predictor).

    o   solve() gives clearer error messages for some incorrect usages.
    (PR#8494 and similar)

    o   The gaussian() family was missing the 'valideta' component
    (which could be needed for the "inverse" link function).

    The starting values supplied by the gaussian family could be
    invalid for the "log" and "inverse" link functions.  This is
    now reported.

    o   data.matrix() did not work correctly on zero-row data frames.
    (PR#8496 and other problems.)

    o   The DSC comments in the files from postscript(onefile=FALSE) now
    label all files as having page 1 of 1, as some other software
    seems to expect that.

    o   The axis labels chosen for logarithmic axis are now less
    likely to be linear and inappropriate (when the range is more
    than 10 and less than 100).  (PR#1235)

    o   Staircase lines (types "s" and "S") are now drawn
    continuously rather than a point at a time and so line types,
    mitring and so on work. (PR#2630)

    o   Calling par(mfg) before doing any plotting resulted in NewPage
    never being called on the device, which in turn resulted in
    incorrect output for postcript() and pdf() devices.  (Reported
    by Marc Schwartz in discussion of the non-bug PR#7820.)

    o   terms.formula needed to add parentheses to formulae with
    terms containing '|'.  (PR#8462)

    o   pbirthday() and qbirthday() now also work for very improbable
    events {those you are typically *not* interested in}.

    o   Only source help files starting with an upper- or lower-case
    letter or digit and extension .Rd or .rd are documented to be
    processed.  This is more liberal in that starting with a digit
    is now also allowed, but rule is now enforced.

    o   nls(algorthm="port") was always taking positive numeric
    differences and so could exceed the upper bounds.

    o   methods:::.asEnvironmentPackage() was not allowing for
    versioned installs.

    o   .find.package() now reports which package(s) it cannot find in
    the case it stops with an error.

    o   The standard Unix-alike version of file.show() gives an
    informative message if it cannot open a file rather than the
    (possibly incorrect) 'NO FILE'.

    o   window() did not allow non-overlapping ranges with
    extend = TRUE.  (PR#8545)

    o   pbinom(size = 0) now returns correct values (not NaN).  (PR#8560)

    o   [dp]binom(x, *) for x < 0 now always returns 0. (PR#8700)
    analogous change in pgeom(), pnbinom() and ppois().

    o   [dqpr]geom and [dpqr]nbinom() now all consistently accept prob = 1
    but not prob = 0.  qgeom(prob=1) now gives the correct values
    (not -1).

    o   INSTALL on Unix-alikes was not loading dependent packages when
    preparing for lazy-loading.

    o   qcauchy(1) now gives +Inf instead of just a very large number.

    o   df(0, f1, *) now properly returns Inf, 1, or 0 for f1 < , = , or > 2.

    o   qbinom(), qnbinom() and qpois() now use a better search and
    normally reach the answer very quickly when it is large
    (instead of being slow or infinite-looping).

    o   pt(x, df) lost accuracy in the far tails (when |x| > 1e154) for
    small df (like df = 0.001 for which such extremes are not unlikely).

    o   dbeta(x, a, b) underflowed internally and incorrectly gave 0
    for very small x and a.

    o   None of the warnings about convergence failures or loss of
    precision in nmath (distribution and special functions) were
    being reported to the R user.

    o   dnt was missing from standalone nmath (under Unix-alikes).

    o   split() now accepts factors with numeric (but not storage mode
    integer) codes.

    o   The utilities such as 'check' now report active version numbers
    again, as SVN 'last changed revision' numbers.

    o   addmargins() did not accept a name for 'FUN', only an expression.

    o   '+' for POSIXt objects now takes the tzone from whichever
    object has it, so date+x is the same as x+date if x is numeric.

    o   mean.default() and var() compute means with an additional pass
    and so are often more accurate, e.g. the variance of a
    constant vector is (almost) always zero and the mean of such a
    vector will be equal to the constant value to machine
    precision.  (PR#1228)

    sum(), prod(), mean(), rowSums() and friends use a long double
    accumulator where available and so may be more accurate.  (This
    is particularly helpful on systems such as Sparc and AMD64
    where long double gives considerably greater exponent range
    and precision than double.)

    o   read.dcf() now gives a warning on malformed lines.

    o   add1.[g]lm now try harder to use the environment of the
    formula in the orginal fit to look for objects such as the
    'data' and 'subset' arguments.

    o   gaussian()$aic was inconsistent with e.g. the lm results from
    AIC() and extractAIC() for weighted fits: it treated the weights
    as case weights and not variance factors.

    o   system() on Unix-alikes ignored non-logical values of 'intern' and
    treated 'intern = NA' as true.

    o   as.table() now produces non-NA rownames when converting a
    matrix of more than 26 rows.  (PR#8652)

    o   Partial sorting used an algorithm that was intended only for a
    few values of 'partial' and so could be far slower than a full
    sort.  It now switches to a barebones full sort for more than
    10 values of 'partial' and uses a more efficient recursive
    implementation for 2...10.

    o   summary.glm() returned an estimate of dispersion of Inf for a
    gaussian glm with zero residual degrees of freedom and then
    treated that as a known value.  It now uses the estimate NaN,
    which is consistent with summary.lm().

    o   Sys.sleep() on Unix-alikes was restricted to about 2147 seconds
    and otherwise might never have returned.  (PR#8678)

    o   is(obj, Cl) could wrongly report TRUE when Cl was a classUnion and
    multiple inheritance was involved.

    o   confint[.lm / .default] used label "100 %" for level = 0.999

    o   Empty entries (i.e., extraneous ",") in NAMESPACE files now give
    a better error message early at parsing time instead of a less
    comprehensible one later at load time.

    o   all.equal(n1, n2) could erroneously return NA when n1, n2
    contained large integers.

    o   anova.mlm() didn't handle multi-df effects properly in the
    single-model case (PR#8679)

    o   anova.mlm() had its colnames mangled by data.frame() (needed
    check.names=FALSE).

    o   summary.glm() gave an NA estimate of dispersion for fits with
    zero weights.  (PR#8720)

    o   qhyper() had too small a tolerance for right-continuity on
    some platforms so was not always an inverse to phyper().

    o   rownames<-.data.frame() and dimnames<-.data.frame() tested the
    length(s) of the replacement value(s) before coercion, which
    can change the length (e.g. for class "POSIXlt").

    o   max() and min() ignored the largest/smallest representable
    integer, as well as Inf/-Inf.  (PR#8731)

    o   write.table() assumed factors had integer codes: it now allows
    malformed factors with numeric codes (and otherwise throws an
    error).

    o   Worked around a Solaris restriction which meant that
    Sys.sleep() was only effective for times of up to one second.

    o   sink(, split=TRUE) now works correctly, but is allowed only on
    platforms that support va_copy or __va_copy.  (PR#8716)

    o   factanal(), prcomp() and princomp() now only check that columns
    in the model frame that will be used are numeric (they
    previously also checked columns which were part of negative terms
    in the formula).

    o   Misuse of $ in apply could corrupt memory.  (PR#8718)

    o   apply() could fail if the function returned NULL (e.g. if
    there was a single row).

    o   registerS3method() failed due to a typo. (It was almost never
    used.)

    o   Registering an S3 method for an S3 generic in another package
    that was converted to an S4 generic in the same package as the
    S3 method, registered the method in the wrong place.

    o   Recall() used lookup for the function in use and so could fail
    if that was an S3 method not on the search path.

    o   Rdconv -t Ssgm failed if it encountered \link[opt]{arg}.

    o   uniroot() did not give a warning (as documented) if it failed
    to converge in 'maxiter' steps.  (PR#8751)

    o   eapply() (and as.list.environment()) did not work for the base
    environment/name space.  (PR#8761)

    o   Added protection in configure against systems for which using
    xmkmf fails to report a C or C++ compiler.

    o   expand.grid() was constructing a data frame 'by hand' and so
    setting integer row.names (which are documented to be
    character).  It now sets character row names, and
    row.names.data.frame() coerces to character.

    o   qbeta() used == on volatile doubles for its convergence test,
    which failed with gcc 3.3.x on ix86 Linux.  We now use a less
    fragile test (and lose a negligible amount of accuracy).

    o   ls.str() was missing inherits=FALSE, and so could have
    reported on an object of the same name but a different mode in
    the enclosure of the given environment.

    o   logLik.nls assumed that sigma^2 had been estimated, but did
    not count this in the 'df' attribute.




    **************************************************
    *                        *
    *          2.2 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 2.2.1 patched


INSTALLATION CHANGES


    o   The macro SOCKLEN_T has been replaced by R_SOCKLEN_T to work
    around a problem with the headers of AIX 5.3.


BUG FIXES

    o   sub(fixed = TRUE) could get wrong the length of the character
    string for elements of the result after the first.

    o   legend() worked out which elements of 'lty' were valid before
    resizing 'lty', and so could fail if 'lty' was a different
    length from 'legend'.

    o   str(<dendrogram>) sometimes used much too many spaces (in 2.2.x).

    o   eigen(eispack=TRUE) accessed areas off the matrix in some
    circumstances (some asymmetric matrices with both complex
    conjugate pair and real eigenvalues).

    o   strptime() in 2.2.1 sometimes did not set $isdst when it was
    previously set.

    o   Another case of infinite influence has been worked around.
    (An addendum to PR#8367.)

    o   qr.coef() worked incorrectly with multiple rhs in the
    LAPACK-using cases. (PR#8476/8)

    o   rbind.data.frame() gave a corrupt data frame if one of the
    named arguments was a zero-row data frame. (PR#8506)

    o   Checks for NULL in the rho argument of the C-level findVar
    function have been added.

    o   The C-level substitute function was handling NULL in its 'rho'
    argument incorrectly.

    o   The code for pgamma() introduced in 2.1.0 failed for large
    values of 'shape' where the previous code was perfectly
    acceptable, despite the claim to be uniformly better.  For
    example, pgamma(0.9e100, 1e100) was NaN.  (PR#8528)

    o   There was no command 'ls' in browser() nor 'next' in debug(),
    despite the documentation (which has been corrected).
    Command 'where' in the browser() no longer changes to
    step-though mode.

    o   factor.scope() could report incorrectly that interaction terms
    were not in the upper scope when such terms in the model and
    the upper scope had different orders for the main effects.
    (Another manifestation of PR#7842.)

    o   The "lm" method of drop1() was giving incorrect results for
    weighted fits (since deviance.lm() was called on a non-"lm" object).

    o   dotchart() was miscalculating the space for the labels in the
    left margin.  (PR#8681)

    o   r <- glm(.....); all.equal(r,r) # now gives TRUE instead of an error

    o   plot.acf() with a multiple time series was sometimes
    miscalculating the 'ylim' value for the plot after the first.
    (PR#8705)



        CHANGES IN R VERSION 2.2.1


USER-VISIBLE CHANGES

    o   options("expressions") has been reduced to 1000: the limit
    of 5000 introduced in 2.1.0 was liable to give crashes from C
    stack overflow.


NEW FEATURES

    o   Use of 'pch' (e.g. in points) in the symbol font 5 is now
    interpreted in the single-byte encoding used by that font.
    Similarly, strwidth now recognizes that font 5 has a different
    encoding from that of the locale.  (These are likely to affect
    the answer only in MBCS locales such as UTF-8.)

    o   The URW font metrics have been updated to versions from late
    2002 which cover more glyphs, including Cyrillic.

    o   New postscript encodings for CP1250 (Windows East European),
    ISO Latin-7 (8859-13, Latvian, Lithuanian and Maori), Cyrillic
    (8859-5), KOI8-R, KOI8-U and CP1251.

    o   configure has more support for the Intel and Portland Group
    compilers on ix86 and x86_64 Linux.

    o   R CMD INSTALL will clean up if interrupted (e.g. by ctrl-C from
    the keyboard).

    o   There is now a comprehensive French translation of the messages,
    thanks to Philippe Grosjean, Frederic Lehobey, Jean Thioulouse
    and Emmanuel Paradis.


DEPRECATED & DEFUNCT

    o   The undocumented use of atan() with two arguments is deprecated:
    instead use atan2() (as documented).

    o   The 'vfont' argument of axis() and mtext() is deprecated
    (it currently warns and does nothing).

    o   The function mauchley.test() is deprecated (was a misspelling)
    and replaced by mauchly.test()


BUG FIXES

    o   The malloc's of AIX and OSF/1 which return NULL for size 0
    are now catered for in src/main/regex.c.

    o   Names of list elements which are missing are now printed as
    $<NA> and not $"NA" (which is how the non-missing name "NA" is
    printed).  (Brought up in discussion of PR#8161.)

    o   help.start() was not linking R.css for use by its front page and
    immediate links (2.2.0 only).

    o   Indexing by character NA matched the name "NA".

    o   The arith-true test used random inputs and did not set the seed, so
    it could fail very occasionally.

    o   arima() with 'fixed' supplied and p=0 for the non-seasonal
    part could give spurious warnings about 'some AR parameters
    were fixed'.

    o   summary.matrix() could give an infinite recursion on some
    classed objects (e.g. those of class "Surv").

    o   The 255th character in an 8-bit character set was not handled
    correctly as a letter on some platforms where C char is
    signed: for example it was printed as \377 and not allowed in
    variable names.  (Spotted by Alexey Shipunov in Russian
    encodings.)

    o   Conversion from POSIXct to POSIXlt is done more accurately
    around the change of DST in years not supported by the OS
    (pre-1970 on Windows and some others, and in the far past or
    future).

    o   chisq.test(cbind(1:0, c(7,16)), simulate.p = TRUE) gave wrong
    P-values on some platforms. (PR#8224)

    o   pdf() was not writing details of the encoding to the file
    correctly.  (Spotted by Alexey Shipunov in Russian encodings.)

    o   image() was failing with an error when plotting a matrix
    of all NA values.  (PR#8228)

    image() could fail if called with add=FALSE (the default) and
    length(x)=1 for either x or y, as it uses the plot coordinates of
    the previous plot (if any).

    o   tools::checkMD5sums was not accepting file names with spaces in.

    o   The plot() method for TukeyHSD() needed updating after adding
    adjusted p-values.  (PR#8229)

    o   read.fwf() did not work for header = TRUE.  (PR#8226)

    o   diag() failed when its argument had NA values in its dimnames.

    o   [g]sub(pcre=TRUE) did not work correctly with \U and \L in a
    UTF-8 locale, even on the example on the help page.

    o   promptMethods() was failing if the "methods" argument was supplied.

    o   is.loaded() now finds Fortran symbols whether or not the
    registration mechanism has been used.

    o   ISODateTime() mistakenly corrected non-existent times (when
    DST was being started) in the current time zone.

    o   Some replacement operations on data frames gave incorrect
    answers, e.g. DF[3:4, "y"] if column "y" did not exist or was
    a matrix.

    o   getGraphicsEvent() would cause memory corruption if passed an
    empty prompt.

    o   qr() and chol() now pivot the colnames of the result when
    pivoting is used.  (PR#8258)

    o   example(points) omitted pch=0, although it was valid and
    said in the text to be illustrated.

    o   plot.default() had an unused 'lab' argument, thereby preventing
    the 'lab' graphics parameter being passed through '...' .

    o   Although polygon(col = NA) was the stated default, specifying
    NA was not equivalent to omitting the argument (but col=NULL
    was equivalent).

    o   Im(-1) was pi. (PR#8272, a side effect from all previous
    versions of R returning the same value for Im and Arg of
    non-complex numbers.)

    o   symbols(fg) defaulted to colour 1, not par("col") as documented.
    It does now defaults to par("col").

    o   par("family") did not check the length of the value (up to 49
    bytes) and so could segfault.

    o   aggregate.ts() did not allow for rounding in frequencies such
    as 1/5.

    o   prcomp(tol=) was not dropping the sdev's corresponding to
    dropped columns.

    o   Subassignment of a vector which increased the length of the
    vector _and_ had the wrong length of replacement could
    occasionally segfault.  (This has been there since at least
    mid 1997.)

    o   The registration of .Fortran symbols was broken: these could
    only be looked up if there were also .Call symbols registered!

    o   R CMD build was incorrectly rejecting the recommended form of
    name for a translation package, 'Translation-ll'.  (PR#8314)

    o   numericDeriv() gave nonsense results unless the variables were
    real, which was not checked.

    o   predict.prcomp() would sometimes give an error when predicting
    a single observation.  (PR#8324)

    o   mapply() could segfault if MoreArgs was not a list.  (PR#8332)

    o   The arith-true test used identical() on floating-point results,
    and this allowed a failure when the relative difference was
    less than .Machine$double.eps but non-zero.

    o   qbinom() was not accepting p = -Inf when log.p = TRUE, although
    it is a legitimate value.

    o   write.csv[2] only accepted logical constants for 'row.names', and
    now accepts variables.

    o   Conversion of .Rd files did not correctly match braces
    enclosing a whole argument, e.g. \eqn{{\bf a}}{a}.

    o   The C function pythag (used if hypot was not available) would
    infinite-loop on systems with effective optimizing compilers.

    o   Writing long formats (more than 1000 bytes) with connections
    that use dummy_vfprintf could fail on some systems.  The limit
    has been changed to 100000 bytes pending a more complete fix in R
    2.3.0.

    o   Making in src/nmath/standalone without making R was not
    making Rmath.h.

    o   Both the R front-end and INSTALL could find the attempted
    temporary directory name already in use on platforms without
    mktemp (and a genuine Bourne shell /bin/sh, not bash).  Now
    both the process ID and a timestamp are used to create the
    directory name.

    o   [dpqr]gamma now return NaN for an invalid 'shape' parameter
    (rather than throw an error), for consistency with other
    distribution functions.

    o   t() now longer drops dimnames 'list(NULL,NULL)' or 'list(NULL)'.

    o   Influence measures such as rstandard() and cooks.distance()
    could return infinite values rather than NaN for a case which
    was fitted exactly.  Similarly, plot.lm() could fail on such
    examples.  plot.lm(which = 5)  had to be modified to only plot
    cases with hat < 1.  (PR#8367)

    lm.influence() was incorrectly reporting 'coefficients' and
    'sigma' as NaN for cases with hat = 1, and on some platforms
    not detecting hat = 1 correctly.

    o   Rmath.h for standalone Rmath was not recording HAVE_WORKING_LOG,
    so R_log was not available on platforms defining it.

    o   HoltWinters() was using a slightly incorrect formula in the C code.

    o   dir.create() could be confused by a trailing slash on the
    path, and by paths containing drives on Windows.

    o   The search for tcl/tkConfig.sh looked in 'lib' before 'lib64'
    directories (and not at all in /usr/local/lib64) and so might
    prefer 32- to 64-bit versions if both are available.

    o   nlminb() used an uninitialized variable unless bounds were supplied,
    and so failed on 64-bit Solaris.



        CHANGES IN R VERSION 2.2.0


USER-VISIBLE CHANGES

    o   plot(<lm object>) uses a new default 'which = 5'
    for the fourth panel when 'which' is not specified.

    o   The SVN revision number will appear after the date in the
    welcome message.  The date shown is now the date of the last
    change to the sources rather than the date the sources were
    prepared.

    o   is.null(expression()) now returns FALSE.  Only NULL gives TRUE
    in is.null().

    o   graphics::xy.coords, xyz.coords and n2mfrow have been moved to
    the grDevices name space (to be available for grid as well).

    graphics::boxplot.stats, contourLines, nclass.*, and chull
    have been moved to the grDevices name space.  The C code
    underlying chull() has been moved to package grDevices.

    o   split(x, f), split<-() and unsplit() now by default split by all
    levels of a factor f, even when some are empty.
    Use split(x, f, drop = TRUE) if you want the old behavior of
    dropping empty levels.  split() and split<-() are S3 generic
    functions with new arguments 'drop' and '...' and all methods now
    should have 'drop' and '...' arguments as well.

    o   The default for 'allowEscapes' in both read.table() and scan()
    has been changed to FALSE.

    o   The default for 'gcFirst' in system.time() is now TRUE.


NEW FEATURES

    o   .Platform has a new component 'path.sep', the separator used
    between paths in environment variables such as PATH and TEXINPUTS.

    o   anova.mlm() now handles the single-model case.

    o   Hexadecimal values are now allowed for as.numeric() and
    as.integer() on all platforms, and as integer constants in R code.

    o   attach() now prints an information message when objects are
    masked on the search path by or from a newly attached database.

    o   axis() now returns 'at' positions.

    o   axis() has a new argument 'hadj' to control horizontal
    adjustment of labels.

    o   axis.Date() and axis.POSIXct() now accept a 'labels' argument
    (contributed by Gavin Simpson).

    o   barplot() now has arguments 'log = ""' and 'add = FALSE'
    (as in barplot2() from package 'gplots').

    o   baseenv() has been added, to return the base environment.  This
    is currently still NULL, but will change in a future release.

    o   boxplot() now responds to supplying 'yaxs' (via bxp()).
    (Wish of PR#8072.)

    o   capabilities() has a new component 'NLS'.

    o   cbind() and rbind() now react to 'deparse.level' = {0,1,2}
    (as in another system not unlike R).

    o   Experimental versions of cbind() and rbind() in methods package,
    based on new generic function cbind2(x,y) and rbind2().  This will
    allow the equivalent of S4 methods for cbind() and rbind() ---
    currently only after an explicit activation call, see ?cbind2.

    o   New functions cdplot() and spineplot() for conditional density
    plots and spine plots or spinograms.  Spine plots are now used
    instead of bar plots for x-y scatterplots where y is a factor.

    o   checkDocFiles() in package 'tools' now checks for bad \usage
    lines (syntactically invalid R code).

    o   The nonparametric variants of cor.test() now behave better in
    the presence of ties. The "spearman" method uses the asymptotic
    approximation in that case, and the "kendall" method likewise,
    but adds a correction for ties (this is not necessary in the
    Spearman case).

    o   The X11 dataentry() now has support for X Input Methods
    (contributed by Ei-ji Nakama).

    o   density() is now an S3 generic where density.default() {former
    density()} has new argument 'weights' for specifying observation
    masses different than the default 1/N -- based on a suggestion and
    code from Adrian Baddeley.

    o   download.packages() now carries on if it encounters a download
    error (e.g. a repository with a corrupt index).

    o   dump() now skips missing objects with a warning rather than
    throw an error.

    o   Added "POSIXlt" methods for duplicated() and unique().

    o   Function encoded_text_to_latex() in package tools translates
    Latin 1,2,9 and UTF-8 encoded character vectors to LaTeX
    escape sequences where appropriate.

    o   encodeString() allows justify = "none" for consistency with
    format.default().  Some argument names have been lengthened
    for clarity.

    o   file(), fifo() and pipe() now (if possible) report a reason
    if they fail to open a connection.

    o   format.default() now has a 'width' argument, and 'justify' can
    now centre character strings.

    format.default() has new arguments 'na.encode' to control
    whether NA character strings are encoded (true by default),
    and 'scientific' to control the use of fixed/scientific
    notation for real/complex numbers.

    How format() works on a list is now documented, and uses
    arguments consistently with their usage on an atomic vector.

    o   format.info() now has a 'digits' argument, and is documented
    to work for all atomic vectors (it used to work for all but
    raw vectors.).

    o   New function glob2rx() for translating `wildcard' aka `globbing'
    to regular expressions.

    o   There is a new function gregexpr() which generalizes regexpr()
    to search for all matches in each of the input strings (not
    just the first match).

    o   [g]sub() now have a 'useBytes' argument like grep() and regexpr().

    o   [g]sub(perl = TRUE) support \L and \U in the replacement.

    o   iconv() has been moved from 'utils' to 'base'.

    o   identify()'s default method has additional arguments 'atpen'
    and 'tolerance' (following S).

    o   KalmanForecast() and KalmanLike() now have an optional argument
    fast=FALSE to prevent their arguments being modified.

    o   Exact p-values are available in ks.test() for the one-sided and
    two-sided one-sample Kolmogorov-Smirnov tests.

    o   labels() now has a method for "dist" objects (replacing that
    for names() which was withdrawn in 2.1.0).

    o   library() now explicitly checks for the existence of
    directories in 'lib.loc': this avoids some warning messages.

    o   loadNamespace(keep.source=) now applies only to that name space
    and not others it might load to satisfy imports: this is now
    consistent with library().

    o   match.arg() has a new argument 'several.ok = FALSE'.

    o   max.col() has a new argument for non-random behavior in the
    case of ties.

    o   memory.profile() now uses the type names returned by typeof()
    and no longer has two unlabelled entries.

    o   methods() now warns if it appears to have been called on a
    non-generic function.

    o   The default mosaicplot() method by default draws grey boxes.

    o   nlminb(), similar to that in S-PLUS, added to package 'stats'.

    o   New algorithm "port" (the nl2sol algorithm available in the
    Port library on netlib) added to the nls() function in the
    'stats' package.

    o   object.size() now supports more types, including external
    pointers and weak references.

    o   options() now returns its result in alphabetical order, and is
    documented more comprehensively and accurately.  (Now all
    options used in base R are documented, including
    platform-specific ones.)

    Some options are now set in the package which makes use of
    them (grDevices, stats or utils) if not already set when the
    package is loaded.

    o   New option("OutDec") to set the decimal point for output conversions.

    o   New option("add.smooth") to add smoothers to a plot, currently
    only used by plot.lm().

    o   pie() has new optional arguments 'clockwise' and 'init.angle'.

    o   plot.lm() has two new plots (for 'which' = 5 or 6), plotting
    residuals or cook distances versus (transformed) leverages - unless
    these are constant.  Further, the new argument 'add.smooth' adds a
    loess smoother to the point plots by default, and 'qqline = TRUE'
    adds a qqline() to the normal plot.
    The default for 'sub.caption' has been improved for long calls.

    o   R.home() has been expanded to return the paths to components
    (which can as from this version be installed elsewhere).

    o   readbin() and writeBin() now support raw vectors as well as
    filenames and connections.

    o   read.dcf() can now read gzipped files.

    o   read.table() now passes 'allowEscapes' to scan().

    o   sample(x, size, prob, replace = TRUE) now uses a faster
    algorithm if there are many reasonably probable values.  (This
    does mean the results will be different from earlier versions
    of R.)  The speedup is modest unless 'x' is very large _and_
    'prob' is very diffuse so that thousands of distinct values
    will be generated with an appreciable frequency.

    o   scatter.smooth() now works a bit more like other plotting
    functions (e.g., accepts a data frame for argument 'x').
    Improvements suggested by Kevin Wright.

    o   signif() on complex numbers now rounds jointly to give the
    requested number of digits in the larger component, not
    independently for each component.

    o   New generic function simulate() in the 'stats' package with
    methods for some classes of fitted models.

    o   smooth.spline() has a new argument 'keep.data' which allows to
    provide residuals() and fitted() methods for smoothing splines.

    o   Attempting source(file, chdir=TRUE) with a URL or connection
    for 'file' now gives a warning and ignores 'chdir'.

    o   source() closes its input file after parsing it rather than
    after executing the commands, as used to happen prior to
    2.1.0.  (This is probably only significant on Windows where
    the file is locked for a much shorter time.)

    o   split(), split<-(), unsplit() now have a new argument 'drop = FALSE',
    by default not dropping empty levels; this is *not* back compatible.

    o   sprintf() now supports asterisk `*' width or precision
    specification (but not both) as well as `*1$' to `*99$'.  Also the
    handling of `%' as conversion specification terminator is now
    left to the system and doesn't affect following specifications.

    o   The plot method for stl() now allows the colour of the range
    bars to be set (default unchanged at "light gray").

    o   Added tclServiceMode() function to the tcltk package to allow
    updating to be suspended.

    o   terms.formula() no longer allows '.' in a formula unless there
    is a (non-empty) 'data' argument or 'allowDotAsName = TRUE' is
    supplied.  We have found several cases where 'data' had not
    been passed down to terms() and so '.' was interpreted as a
    single variable leading to incorrect results.

    o   New functions trans3d(), the 3D -> 2D utility from persp()'s
    example, and extendrange(), both in package 'grDevices'.

    o   TukeyHSD() now returns p-values adjusted for multiple comparisons
    (based on a patch contributed by Fernando Henrique Ferraz P. da Rosa).

    o   New functions URLencode() and URLdecode(), particularly for use
    with file:// URLs.  These are used by e.g. browse.env(),
    download.file(), download.packages() and various help() print
    methods.

    o   Functions utf8ToInt() and intToUtf8() to work with UTF-8
    encoded character strings (irrespective of locale or OS-level
    UTF-8 support).

    o   [dqp]wilcox and wilcox.test work better with one very large sample
    size and an extreme first argument.

    o   write() has a new argument 'sep'.

    o   write.csv[2] now also support row.names = FALSE.



    o   The specification of the substitutions done when processing
    Renviron files is more liberal: see ?Startup.  It now
    accepts forms like R_LIBS=${HOME}/Rlibrary:${WORKGRP}/R/lib .

    o   Added recommendation that packages have an overview man page
    <pkg>-package.Rd, and the promptPackage() function to create a
    skeleton version.

    o   Replacement indexing of a data frame by a logical matrix index
    containing NAs is allowed in a few more cases, in particular
    always when the replacement value has length one.

    o   Conversion of .Rd files to latex now handles encoding more
    comprehensively, including some support for UTF-8.

    o   The internal regex code has been upgraded to glibc-2.3.5.
    Apart from a number of bug fixes, this should be somewhat
    faster, especially in UTF-8 locales.

    o   PCRE has been updated to version 6.2.

    o   zlib has been updated to version 1.2.3.

    o   bzip2 has been updated to version 1.0.3.

    o   Complex arithmetic is now done by C99 complex types where
    supported.  This is likely to boost performance, but is
    subject to the accuracy with which it has been implemented.

    o   The printing of complex numbers has changed, handling numbers
    as a whole rather than in two parts.  So both real and
    imaginary parts are shown to the same accuracy, with the
    'digits' parameter referring to the accuracy of the larger
    component, and both components are shown in fixed or
    scientific notation (unless one is entirely zero when it is
    always shown in fixed notation).

    o   Error messages from .C() and .Fortran(), and from parsing errors,
    are now more informative.

    o   The date and date-time functions work better with dates more
    than 5000 years away from 1970-01-01 (by making dubious
    assumptions about the calendar in use).

    o   There is now a traditional Chinese translation, and a much more
    extensive Russian translation.


DEPRECATED & DEFUNCT

    o   Capability "IEEE754" is defunct.

    o   loadURL() is defunct: use load(url()).

    o   delay() is defunct: use delayedAssign() instead.

    o   The 'CRAN' argument to update.packages(), old.packages(),
    new.packages(), download.packages() and install.packages() is
    defunct in favour of 'repos'.

    o   write.table0() is deprecated in favour of the much faster
    write.table().

    o   format.char() is deprecated in favour of format.default().

    o   R_HOME/etc/Rprofile is no longer looked for if
    R_HOME/etc/Rprofile.site does not exist.  (This has been
    undocumented since R 1.4.0.)

    o   CRAN.packages() is deprecated in favour of available.packages().

    o   Rd.sty no longer processes pre-2.0.0 conversions containing \Link.

    o   The stubs for the defunct device GNOME/gnome have been removed.

    o   print.matrix() (which has been identical to print.default since
    R 1.7.0) has been removed.


INSTALLATION

    o   LDFLAGS now defaults to -L/usr/local/lib64 on most Linux
    64-bit OSes (but not ia64).  The use of lib/lib64 can be
    overridden by the new variable LIBnn.

    o   The default installation directory is now ${prefix}/${LIBnn}/R,
    /usr/local/lib64/R on most 64-bit Linux OSes and /usr/local/lib/R
    elsewhere.

    o   The places where the doc, include and share directory trees are
    installed can be specified independently: see the R-admin manual.

    o   We now test for wctrans_t, as apparently some broken OSes have
    wctrans but not wctrans_t (which is required by the relevant
    standards) .

    o   Any external BLAS found is now tested to see if the complex
    routine zdotu works correctly: this provides a compatibility
    test of compiler return conventions.

    o   Installation without NLS is now cleaner, and does not install
    any message catalogues.

    o   src/modules/lapack/dlamc.f is now compiled with -ffloat-store
    if f2c/gcc are used, as well as if g77 is used.

    o   All the Fortran code has been checked to be fully F77
    compliant so there are no longer any warnings from F95
    compilers such as gfortran.

    o   The (not-recommended) options --with-system-zlib,
    --with-system-bzlib and -with-system-pcre now have 'system' in
    the name.

    o   If a Java runtime environment is detected at configure time
    its library path is appended to LD_LIBRARY_PATH or equivalent.
    New Java-related variables JAVA_HOME (path to JRE/JDK), JAVA_PROG
    (path to Java interpreter), JAVA_LD_PATH (Java library path)
    and JAVA_LIBS (flags to link against JNI) are made available
    in Makeconf.

    o   Ei-ji Nakama was contributed a patch for FPU control with the
    Intel compilers on ix86 Linux.


MAC OS X INSTALLATION

    o   --with-blas="-framework vecLib" --with-lapack and
    --with-aqua are now the default configure options.

    o   The default framework version name was changed to not contain
    the patch level (i.e. it is now 2.2 instead of 2.2.0).  Also
    it can be overridden at configure time by setting FW_VERSION
    to the desired name.

    o   The Rmath stand-alone library is now correctly installed inside
    the R.framework if R was configured as a framework.  In
    addition, make install-Rmath-framework will install a
    stand-alone Rmath framework in /Library/Frameworks (unless
    overridden by RMATH_FRAMEWORK_DIR specifying full framework
    path and name including the .framework extension).


PACKAGE INSTALLATION

    o   The encoding for a packages' 00Index.html is chosen from the
    Encoding: field (if any) of the DESCRIPTION file and from the
    \encoding{} fields of any Rd files with non-ASCII titles.
    If there are conflicts, first-found wins with a warning.

    o   R_HOME/doc/html/packages.html is now remade by R not Perl code.
    This may result in small changes in layout and a change in
    encoding (to UTF-8 where supported).

    o   The return value of new.packages() is now updated for any
    packages which may be installed.

    o   available.packages() will read a compressed PACKAGES.gz file in
    preference to PACKAGES if available on the repository: this
    will reduce considerably the download time on a dialup connection.

    The downloaded information about a repository is cached for the
    current R session.

    o   The information about library trees found by
    installed.packages() is cached for the current session, and
    updated only if the modification date of the top-level
    directory has been changed.

    o   A data index is now installed for a package with a 'data' dir
    but no 'man' dir (even though it will have undocumented data objects).

    o   contrib.url path for type="mac.binary" has changed from
    bin/macosx/<version> to bin/macosx/<arch>/contrib/<version>
    where <arch> corresponds to R.version$arch


UTILITIES

    o   checkFF() used by R CMD check has since R 2.0.0 not reported
    missing PACKAGE arguments when testing installed packages with
    name spaces.  It now

    - treats installed and source packages in the same way.

    - reports missing arguments unless they are in a function in
      the name space with a useDynLib declaration (as the
      appropriate DLL for such calls can be searched for).

    o   Rd2dvi sets the encoding(s) used appropriately.  If UTF-8
    encoding is used, latex >= 2003/12/01 is required.

    o   codoc() allows help files named pkg_name-defunct.Rd to have
    undocumented arguments (and not just base-defunct.Rd).


C-LEVEL FACILITIES

    o   C function massdist() {called from density()} has new argument
    'xmass' (= weights).

    o   Raw vectors passed to .C() are now passed as unsigned char *
    rather than as SEXPs.  (Wish of Keith Frost, PR#7853)

    o   The search for symbols in a .C/.Call/... call without a
    package argument now searches for an enclosing name space and
    so finds functions defined within functions in a name space.

    o   R_max_col() has new (5th) argument '*ties_meth' allowing
    non-random behavior in the case of ties.

    o   The header files have been rationalized: the BLAS routine
    LSAME is now declared in BLAS.h not Linpack.h, Applic.h no
    longer duplicates routines from Linpack.h, and Applic.h is
    divided into API and non-API sections.

    o   memory.c has been instrumented so that Valgrind can track R's
    internal memory management.  To use this, configure using
           --with-valgrind-instrumentation=level
    where level is 1 or 2.  Both levels will find more bugs with
    gctorture(TRUE).  Level 2 makes Valgrind run extremely slowly.

    o   Some support for raw vectors has been added to Rdefines.h.

    o   R_BaseEnv has been added, to refer to the base environment.
    This is currently equal to R_NilValue, but it will change in
    a future release.


BUG FIXES

    o   %/% has been adjusted to make x == (x %% y) + y * ( x %/% y )
    more likely in cases when extended-precision registers were
    interfering.

    o   Operations on POSIXct objects (such as seq(), max() and
    subsetting) try harder to preserve time zones and warn if
    inconsistent time zones are used.

    o   as.function.default() no longer asks for a bug report when
    given an invalid body. (PR#1880, PR#7535, PR#7702)

    o   Hershey fonts and grid output (and therefore lattice output)
    now rescale correctly in fit-to-window resizing on a Windows
    graphics device.  Line widths also scale now.

    o   Plotmath has more support for multibyte characters (contributed
    by Ei-ji Nakama).

    o   The X11() device now hints the window manager so that decorations
    appear reliably under e.g. the GNOME WM  (contributed
    by Ei-ji Nakama).

    o   Subsetting a matrix or an array as a vector used to attempt to
    use the row names to name the result, even though the
    array might be longer than the row names.  Now this is only
    done for 1D arrays when it is done in all cases, even matrix
    indexing.  (Tidies up after the fix to PR#937.)

    o   Constants in mathlib are declared 'const static double' to
    avoid performance issues with the Intel Itanium compiler.

    o   The parser checks the format of numeric constants more
    thoroughly so for example '123E-' is no longer valid.

    o   contourLines() no longer requires an open device (used to
    start a device unnecessarily).  Fix suggested by Barry Rowlingson.

    o   capabilities() used partial matching but was not documented
    to: it no longer does so.

    o   kernel(1,0) printed wrongly;  kernel(<name-string>, *) now returns
    a named kernel in all cases; plot(kernel(.),..) is more flexible.

    o   qgamma(1,s) didn't give +Inf for some s.

    o   installed.packages() and download.packages() now always
    return a matrix as documented, possibly with 0 rows (rather than
    a 0-length character vector or NULL).

    o   Arithmetic operations on data frames no longer coerce the
    names to syntatically valid names.

    o   Units are now properly recycled in grid layouts
    when 'widths' or 'heights' are shorter than the number of
    columns or rows (PR#8014).

    o   DF <- data.frame(A=1:2, B=3:4); DF[1, 1:3] <- NULL gave a wrong
    error message.

    o   spline()/spinefun()'s C code had a memory access buglet which
    never lead to incorrect results.  (PR#8030)

    o   sum() was promoting logical arguments to double not integer
    (as min() and other members of its group do).

    o   loess() had a bug causing it to occasionally miscalculate
    standard errors (PR#7956).  Reported by Benjamin Tyner, fixed
    by Berwin Turlach.

    o   library(keep.source=) was ignored if the package had a
    name space (the setting of options("keep.source.pkgs") was
    always used).

    o   hist.POSIXct() and hist.Date() now respect par("xaxt").

    o   The 'vfont' argument was not supported correctly in title(),
    mtext(), and axis().  The 'vfont' argument is superseded by
    the par(family=) approach introduced in 2.0.0.  This bug-fix
    just updates the warning messages and documentation to
    properly reflect the new order of things.

    o   The C-level function PrintGenericVector could overflow if
    asked to print a length-1 character vector of several thousand
    characters.  This could happen when printing a list matrix,
    and was fatal up to 2.1.1 and silently truncated in 2.1.1 patched.

    o   What happened for proc.time() and system.time() on
    (Unix-alike) systems which do not support timing was
    incorrectly documented.  (They both exist but throw an error.)
    Further, systen.time() would give an error in its on.exit
    expression.

    o   weighted.residuals() now does sensible things for glm() fits:
    in particular it now agrees with an lm() fit for a Gaussian glm()
    fit.  (PR#7961).

    o   The 'lm' and 'glm' methods for add1() took the weights and
    offset from the original fit, and so gave errors in the
    (dubious) usage where the upper scope resulted in a smaller
    number of cases to fit (e.g. by omitting missing values in new
    variables).  (PR#8049)

    o   demo() had a 'device' argument that did nothing (although it
    was documented to): it has been removed.

    o   Setting new levels on a factor dropped all existing
    attributes, including class "ordered".

    o   format.default(justify="none") now by default converts NA
    character strings, as the other values always did.

    o   format.info() often gave a different field width from format()
    for character vectors (e.g. including missing values or
    non-printable characters).

    o   axis() now ensures that if 'labels' are supplied as character
    strings or expressions then 'at' is also supplied (since the
    calculated value for 'at' can change under resizing).

    o   Defining S4 methods for "[" had resulted in changed behavior of S3
    dispatch in a very rare case which no longer happens.

    o   Fixed segfault when PostScript font loading fails, e.g., when
    R is unable to find afm files (reported by Ivo Welch).

    o   R CMD BATCH <file> now also works when <file> does not end in a
    newline on Unix-alike platforms.

    o   terms.formula() got confused if the 'data' argument was a list with
    non-syntactic names.

    o   prompt() and hence package.skeleton() now produce *.Rd files that
    give no errors (but warnings) when not edited, much more often.

    o   promptClass() and promptMethods() now also escape "%" e.g. in '%*%'
    and the latter gives a message about the file written.

    o   wilcox.test() now warns when conf.level is set higher than
    achievable, preventing errors (PR#3666) and incorrect answers
    with extremely small sample sizes.

    o   The default (protection pointer) stack size (the default for
    '--max-ppsize') has been increased from 10000 to 50000 in order to
    match the increased default options("expressions") (in R 2.1.0).

    o   The R front-end was expecting --gui=tk not Tk as documented,
    and rejecting --gui=X11.

    o   Rdconv -t latex protected only the first << and >> in a chunk
    against conversion to guillemets.

    o   callNextMethod() and callGeneric() have fixes related to
    handling arguments.

    o   ls.diag() now works for fits with missing data. (PR#8139)

    o   window.default() had an incorrect tolerance and so sometimes
    created too short a series if 'start' or 'end' were zero.

    o   Some (fairly pointless) cases of reshape left a
    temporary id variable in the result (PR#8152)

    o   R CMD build used 'tar xhf' which is invalid on FreeBSD systems
    (and followed tar chf, so there could be no symbolic links in
    the tarball).

    o   Subassignment of length zero vectors to NULL gave garbage
    answers. (PR#8157)

    o   Automatic coercion of raw vectors to lists was missing, so for a
    list (or data frame) z, z[["a"]] <- raw_vector did not work
    and now does.  This also affected DF$a <- raw_vector for a
    data frame DF.

    o   The internal code for commandArgs() was missing PROTECTs.

    o   The width for strwrap() was used as one less than specified.

    o   R CMD INSTALL was not cleaning up after an unsuccessful
    install of a non-bundle which was not already installed.




    **************************************************
    *                        *
    *          2.1 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 2.1.1 patched


BUG FIXES

    o   runmed(x, k = -1) now gives an error instead of a seg.fault.

    o   File creation errors in pdf(), postscript(), xfig() resulted
    in a pointer being freed twice.  (Reported by Matt McCall)

    o   model.matrix(~ .^2, data=foo) now works as most people would
    expect (it used to expand '.' after using the a^2 = a rule).

    o   ftable() and xtabs() had a check for interactions that would
    not work correctly with '.' in the formula.

    o   The formula method for pairs() was ignoring the na.action argument.

    o   scan() with the default separator (only) was stripping backslashes
    inside quoted string inputs even if allowEscapes = FALSE.

    o   The "col" parameter to pairs() is now treated consistently with
    plot():  it affects the data, not the axes.  (Patch submitted
    by Olaf Mersmann)

    o   Sweave failed for \Sexpr{character(0)}.  Sweave-test-1.Rnw
    contained a print() statement that is no longer needed.

    o   Typo meant that R_alloc was limited to 2^31 not 2^34 on 64-bit
    builds of R.

    o   pgamma(Inf, shape) did not terminate for shape = 1.1 and some
    other values (but not all) (PR#8001).  This affected
    pchisq(Inf, df1) and pf(Inf, df1, Inf) for some values of df1.

    o   regexpr("[a-z]", ...) could cause a buffer overrun with multi-byte
    character sets, leading to random errors later.

    o   Installing a package with a 'data' directory which contains
    files but those files generate no objects (e.g. the BioC
    package makecdfenv) created an incorrectly formatted data
    index that data(package="pkg") could not read.

    o   density(1/(0:2)) now works again (PR#8033).

    o   make.names() was not respecting the allow_=FALSE argument.

    o   Arg(-1) now gives pi, not 0; Arg(0i + -1) always worked.

    o   --enable-linux-lfs had been broken at 2.1.0.

    o   Printing was not allowing for double-width characters in its
    layout.

    o   axTicks() is now also correct for reverse axis.  (PR#7973)

    o   signif() rounded some numnbers near 1e-308 to the wrong
    number of places (this showed up in the print-tests.R), and made
    unnecessary rounding errors on some platforms on e.g. signif(18000, 3).

    o   window() was sometimes failing incorrectly due to representation
    errors when the new and old deltat were not both integers
    (e.g. multiples of 0.1).

    o   The 'lm' methods for add1() and drop1() ignored offsets (which
    were added for lm() after they written).  (PR#8049)

    o   atan2(0+1i, 0+0i) was incorrectly NA (from a typo in complex.c).

    o   The POSIXct method for as.Date() was rounding part days
    before 1970-01-01 upwards rather than discarding them.

    o   qpois() was using an incorrect starting point and so could be
    unnecessarily slow for large lambda.  (PR#8058)

    o   Formatting of complex numbers with nsmall > 0 could be
    incorrect (and was in print-tests.Rout) because of a typo.

    o   The test for new levels when predicting using model.frame()
    sometimes reported levels that were not actually used.

    o   order(c("5","6",NA,"4",NA), na.last=FALSE) was incorrect.

    o   Coercion to raw could give spurious messages about discarding
    imaginary parts.

    Coercion of a list to raw was behaving inconsistently for
    out-of-range values (and not warning).

    o   cor.test(method = "spearman") gave NA p-values for _very_ long
    vectors. (PR#8087)

    o   Switching to the Mersenne-Twister RNG could cause a segfault on
    first use.

    A user-supplied RNG without user_unif_{nseed,seedloc} was
    being re-initialized at each call.

    o   Reading very short tables from stdin() with read.table would
    fail because of a typo.




        CHANGES IN R VERSION 2.1.1


NEW FEATURES

    o   bug.report() now reports the locale in use.

    o   upgrade.packageStatus() allows user input "c" to cancel the
    upgrade, just as update.packages() does.

    o   glm() now accepts 1D arrays (e.g. tables) as a response,
    dropping them to a vector whilst preserving names.

    o   df() with one infinite df now works (to match pf()).

    o   Added tclServiceMode() function to the tcltk package to allow
    updating to be suspended.

    o   The Encoding: field of a DESCRIPTION file is now documented,
    and used by packageDescription() and library(help=).

    o   There has been progress on translations: existing translations
    have been revised and expanded, and French and Korean have
    been added.

    The Windows installer supports a wide range of languages for
    installation.


BUG FIXES

    o   lm(qr=FALSE) now works.

    o   predict.glm() not longer loses names for "response" predictions.
    (PR#7792)

    o   Typo in menu(graphics=TRUE) meant it failed on Unix if tcltk
    was not available.

    o   When names.dist() was removed, the result of cmdscale() lost
    its rownames.  The example also lost the labels.

    o   R CMD check assumed 'tar' was GNU tar and so supported -z.

    o   read.table() was not handing escaped quotes inside quoted fields
    in the first five lines of the file.  (PR#7789)

    It was also not handling correctly EOF in the first five lines
    when reading from stdin().  (PR#7772)

    o   'make uninstall' was incomplete.

    o   make.packages.html() called by help.start() was failing if
    there were installed packages with help titles invalid in the
    current locale.

    o   printCoefmat(signif.legend = FALSE) was non-functional. (PR#7802)

    o   Some as.date.frame() methods failed because the expression
    deparsed into multiple lines. (PR#7808)

    o   setRepositories() had a typo. (PR#7810)

    o   Printing arrays/data frames with multibyte characters in the
    column labels was sometimes misaligned or using excessive space.
    (PR#7803)

    o   The Tcl/Tk console did not support multibyte characters.

    o   as.POSIXlt() could give infinite recursion if passed a corrupt
    "POSIXct" object (generated by an incorrect call to c.POSIXct,
    PR#7826).

    o   update.packages() was not passing 'type' correctly to
    install.packages().

    o   Printing the result of an unbalanced model.tables() call
    sometimes got confused if terms() had rearranged interaction
    terms.  (PR#7829)

    o   .Platform$pkgType was wrong on the CRAN MacOS X build, and
    .install.macbinary() was missing.

    o   as.personList() as used by citation() got confused by names
    containing "and".  (PR#7797)

    o   Subscripting an array by a matrix containing zero or negative
    values or the wrong number of columns was not handled
    consistently. (PR#7824)

    o   select.list(multiple=TRUE) now detects and tries again for invalid
    text input.

    o   add1.[g]lm could give strange results with interaction terms
    when the model and the upper scope had different orders for
    the main effects.  (PR#7842)

    o   A bug had sneaked into the anova.mlmlist() code, affecting the
    Greenhouse-Geisser epsilon. Code wrongly assumed a matrix to
    be symmetric. (Thanks to Bela Bauer.)

    o   anova.mlmlist() and mauchley.test() are now more tolerant to rank
    deficiency in the M and X matrices (also when they are implicitly
    generated via model.matrix()).

    o   anova.mlm had a scoping issue (PR#7898)

    o   pf() with infinite df is allowed again.  It is now more accurate
    for extreme ratios of dfs, especially when there is a
    non-centrality parameter.

    o   df() was inaccurate for large df (1e16 or greater).

    o   dt() was inaccurate for large df (1e9 or greater) with a
    non-centrality parameter.

    o   runmed(*, algorithm="Turlach") seg.faulted in rare cases.

    o   strwrap() now makes a reasonable job of text that is invalid in the
    current locale.

    o   Reading with encoding "UCS-2LE" will remove any Byte Order
    Mark, as most implementations of iconv fail to handle BOMs
    (which are present in 'Windows Unicode' files).

    o   unique() for a list was incorrectly reporting `unimplemented'.

    o   The parser's contextstack was not protected against overflow,
    e.g. more than 50 unmatched '('.  (PR#7859)

    o   source(file, chdir = TRUE) was not checking that 'file' was a
    filepath (rather than a URL).  For 2.1.0 only, it did not work
    even if 'file' was a filepath.

    o   Hershey fonts were being sized based on pixels not points so came
    out too small on devices where pixels were noticeably different
    from points (e.g., win.printer() and high-resolution screens).

    Fix means that default size of Hershey fonts may be slightly
    different, for example, smaller by default on PostScript and PDF.

    o   The branch cuts in the complex versions of the inverse
    trigonometric and hyperbolic functions were non-standard.
    (PR#7871)

    o   truncate() on file() connections was limited to files < 2Gb.  It
    now works for larger files at least on 64-bit OSes and others
    where ftruncate supports such files.  (Related to PR#7879)

    o   proj.aovlist() did not work correctly on objects fitted from a
    data frame with row names.

    o   The coding standards recommendations had
    nuke-trailing-whitespace where newer versions of ESS need
    ess-nuke-trailing-whitespace.  (PR#7888)

    o   package.skeleton() missed the first newline in the DESCRIPTION file.

    o   pbirthday() reported p = 1 too often when coincident > 2.

    o   plot(1:3, exp(1:3), log = "y", ylim = c(30,1)) {reversed
    log-scale axis} now works, based on Uwe Ligges' suggestions.
    (PR#7894)

    o   install.packages() was aborting when a package in a bundle was
    chosen from a menu.  It failed if more than one package in a
    bundle was chosen from the command line.

    o   qcauchy() suffered from underflow in the extreme tails.  (PR#7902)

    o   Printing of raw matrices/arrays was not implemented. (PR#7912)

    o   getCallingDLL()'s default first argument did not correspond to
    its description and has been changed.  The mismatch caused
    symbols in .C/.Call/.Fortran calls without a PACKAGE= argument
    to be potentially looked up in the wrong name space.

    o   Binary save() of raw vectors was not working correctly on
    big-endian platforms.  (PR#7812)

    o   as.Date.factor() now accepts a format argument.

    o   Workaround added for FreeBSD which does not have alloca.h _and_
    does not allow alloca() to be declared.

    o   identify() now respects 'cex'.  (PR#660)

    Warnings from identify() are now printed immediately even on
    consoles with delayed printing.



        CHANGES IN R VERSION 2.1.0


USER-VISIBLE CHANGES

    o   box plots {by boxplot() or bxp()} now have the median line three
    times the normal line width in order to distinguish it from the
    quartile ones.

    o   Unix-alike versions of R can now be used in UTF-8 locales on
    suitably equipped OSes.  See the internationalization section
    below.

    o   The meaning of 'encoding' for a connection has changed: See the
    internationalization section below.

    o   There has been some rationalization of the format of
    warning/error messages, to make them easier to translate.
    Generally names of functions and arguments are single-quoted,
    and classes double-quoted.

    o   Reading text files with embedded "\" (as in Windows file names)
    may now need to use scan(* , allowEscapes = FALSE), see also below.


NEW FEATURES

    o   %% now warns if its accuracy is likely to be affected by lack
    of precision (as in 1e18 %% 11, the unrealistic expectation of
    PR#7409), and tries harder to return a value in range when it is.

    o   abbreviate() now warns if used with non-ASCII chars, as the
    algorithm is designed for English words.

    o   The default methods for add1() and drop1() check for changes
    in the number of cases in use.

    The "lm" and "glm" methods for add1() quoted the <none> model
    on the original fitted values when using (with a warning) a
    smaller set of cases for the expanded models.

    o   Added alarm() function to generate a bell or beep or visual alert.

    o   all/any() now attempt to coerce their arguments to logical, as
    documented in the Blue Book.  This means e.g. any(list()) works.

    o   New functions for multivariate linear models: anova.mlm(),
    SSD(), estVar(), mauchley.test() (for sphericity).

    vcov() now does something more sensible for "mlm" class objects.

    o   as.data.frame.table() has a new argument 'responseName'
    (contributed by Bill Venables).

    o   as.dist() and cophenetic() are now generic, and the latter has a
    new method for objects of class "dendrogram".

    o   as.ts() is now generic.

    o   binomial() has a new "cauchit" link (suggested by Roger Koenker).

    o   chisq.test() has a new argument 'rescale.p'.  It is now possible
    to simulate (slowly) the P value also in the 1D case
    (contributed by Rolf Turner).

    o   choose(n,k) and lchoose(.) now also work for arbitrary (real) n
    in accordance with the general binomial theorem.  choose(*,k)
    is more accurate (and faster) for small k.

    o   Added colorRamp() and colorRampPalette() functions for color
    interpolation.

    o   colSums()/rowSums() now allow arrays with a zero-length extent
    (requested by PR#7775).

    o   confint() has stub methods for classes "glm" and "nls" that
    invoke those in package MASS.  This avoids using the "lm"
    method for "glm" objects if MASS is not attached.

    confint() has a default method using asymptotic normality.

    o   contr.SAS() has been moved from the 'nlme' package to the
    'stats' package.

    o   New function convertColors() maps between color spaces.
    colorRamp() uses it.

    o   The cov() function in the non-Pearson cases now ranks data after
    removal of missing values, not before.  The pairwise-complete
    method should now be consistent with cor.test.  (Code
    contributed by Shigenobu Aoki.)

    o   Added delayedAssign() function to replace delay(), which is now
    deprecated.

    o   dir.create() has a new argument 'recursive' serving the same
    purpose as Unix's mkdir -p.

    o   do.call() now takes either a function or a character string as
    its first argument.  The supplied arguments can optionally be
    quoted.

    o   duplicated() and unique() now accept "list" objects, but are
    fast only for simple list objects.

    o   ecdf() now has jumps of the correct size (a multiple of 1/n)
    if there are ties.  (Wished by PR#7292).

    o   eff.aovlist() assumed orthogonal contrasts for any term
    with more than one degree of freedom: this is now documented
    and checked for.  Where each term only occurs in only one
    stratum the efficiencies are all one: this is detected and
    orthogonal contrasts are not required.

    o   New function encodeString() to encode character strings in the
    same way that printing does.

    o   file("clipboard") now work for reading the primary selection on
    Unix-alikes with an active X11 display.  (It has long worked
    for reading and writing under Windows.)  The secondary
    selection can also be read: see ?file.

    file() now allows mode "w+b" as well as "w+".

    o   file.append() has been tuned, including for the case of
    appending many files to a single file.

    o   Functions flush.console() and select.list() are now available
    on all platforms.  There is a Tcl/Tk-based version of
    select.list() called tk_select.list() in package tcltk.

    o   gc() now reports maximum as well as current memory use.

    o   A new function getGraphicsEvent() has been added which will allow
    mouse or keyboard input from a graphics device. (NB: currently
    only the Windows screen device supports this function.  This
    should improve before the 2.1.0 release.)

    o   New functions gray.colors()/grey.colors() for gray color
    palettes.

    o   grep(), gsub(), sub() and regexpr() now always attempt to
    coerce their 'pattern', 'x', 'replacement' and 'text'
    arguments to character.  Previously this was undocumented but
    done by [g]sub() and regexpr() for some values of their
    other arguments.  (Wish of PR#7742.)

    o   gsub/sub() have a new 'fixed' method.

    o   New function hcl() for creating colors for a given hue,
    chroma and luminance (i.e. perceptual hsv).

    o   isTRUE() convenience function to be used for programming.

    o   kmeans() now returns an object of class "kmeans" which has a
    print() method.

    Two alternative algorithms have been implemented.

    If the number of centres is supplied, it has a new option of
    multiple random starts.

    o   The limits on the grid size in layout() are now documented, and
    have been raised somewhat by using more efficient internal
    structures.

    o   legend() now accepts positioning by keyword, e.g. "topleft",
    and can put a title within the legend.  (Suggested by Elizabeth
    Purdom in PR#7400.)

    o   mahalanobis() now has a '...' argument which is passed to solve()
    for computing the inverse of the covariance matrix, this replaces
    the former 'tol.inv' argument.

    o   menu() uses a multi-column layout if possible for more than
    10 choices.

    menu(graphics = TRUE) is implemented on most platforms via
    select.list() or tk_select.list().

    o   New function message() in 'base' for generating "simple"
    diagnostic messages, replacing such a function in the
    'methods' package.

    o   na.contiguous() is now (S3) generic with first argument
    renamed to 'object'.

    o   New function normalizePath() to find canonical paths (and on
    Windows, canonical names of components).

    o   The default in options("expressions") has been increased to 5000,
    and the maximal settable value to 500000.

    o   p.adjust() has a new method "BY".

    o   pbeta() now uses a different algorithm for large values of at
    least one of the shape parameters, which is much faster and is
    accurate and reliable for very large values.  (This affects
    pbinom(), pf(), qbeta() and other functions using pbeta at C
    level.)

    o   pch="." now by default produces a rectangle at least 0.01" per
    side on high-resolution devices.  (It used to be one-pixel
    square even on high-resolution screens and Windows printers,
    but 1/72" on postscript() and pdf() devices.)  Additionally,
    the size is now scalable by 'cex'; see ?points and note that
    the details are subject to change.

    o   pdf() now responds to the 'paper' and 'pagecentre'
    arguments.  The default value of 'paper' is "special"
    for backward-compatibility (this is different from the
    default for postscript()).

    o   plot.data.frame() tries harder to produce sensible plots for
    non-numeric data frames with one or two columns.

    o   The predict() methods for "prcomp" and "princomp" now match
    the columns of 'newdata' to the original fit using column
    names if these are available.

    o   New function recordGraphics() to encapsulate calculations
    and graphics output together on graphics engine display list.
    To be used with care.

    o   New function RSiteSearch() to query R-related resources
    on-line (contributed by Jonathan Baron and Andy Liaw).

    o   scan() arranges to share storage of duplicated character
    strings read in: this can dramatically reduce the memory
    requirements for large character vectors which will
    subsequently be turned into factors with relatively few
    levels.  For a million items this halved the time and reduced
    storage by a factor of 20.

    scan() has a new argument 'allowEscapes' (default TRUE) that
    controls when C-style escapes in the input are interpreted.
    Previously only \n and \r were interpreted, and then only
    within quoted strings when no separator was supplied.

    scan() used on an open connection now pushes back on the
    connection its private `ungetc' and so is safer to use to
    read partial lines.

    o   scatter.smooth() and loess.smooth() now handle missing values
    in their inputs.

    o   seq.Date() and seq.POSIXt() now allow 'to' to be before 'from'
    if 'by' is negative.

    o   sprintf() has been enhanced to allow the POSIX/XSI specifiers
    like "%2$6d", and also accepts "%x" and "%X".

    sprintf() does limited coercion of its arguments.

    sprintf() accepts vector arguments and operates on them in
    parallel (after re-cycling if needed).

    o   New function strtrim() to trim character vectors to a display
    width, allowing for double-width characters in multi-byte
    character sets.

    o   subset() now has a method for matrices, similar to that for
    data frames.

    o   Faster algorithm in summaryRprof().

    o   sunflowerplot() has new arguments 'col' and 'bg'.

    o   sys.function() now has argument 'which' (as has long been
    presaged on its help page).

    o   Sys.setlocale("LC_ALL", ) now only sets the locale categories
    which R uses, and Sys.setlocale("LC_NUMERIC", ) now gives a
    warning (as it can cause R to malfunction).

    o   unclass() is no longer allowed for environments and external
    pointers (since these cannot be copied and so unclass() was
    destructive of its argument).  You can still change the
    "class" attribute.

    o   File-name matching is no longer case-insensitive with unz()
    connections, even on Windows.

    o   New argument 'immediate.' to warning() to send an immediate
    warning.

    o   New convenience wrappers write.csv() and write.csv2().

    o   There is a new version for write.table() which is implemented in C.
    For simple matrices and data frames this is several times
    faster than before, and uses negligible memory compared to the
    object size.

    The old version (which no longer coerces a matrix to a data
    frame and then back to a matrix) is available for now as
    write.table0().

    o   The functions xinch(), yinch(), and xyinch() have been moved
    from package 'grDevices' into package 'graphics'.


    o   Plotmath now allows underline in expressions.  (PR#7286,
    contributed by Uwe Ligges.)

    o   BATCH on Unix no longer sets --gui="none" as the X11 module
    is only loaded if needed.

    o   The X11 module (and the hence X11(), jpeg() and png() devices
    and the X-based dataentry editor) is now in principle
    available under all Unix GUIs except --gui="none", and this is
    reflected in capabilities().

    capabilities("X11") determines if an X server can be accessed,
    and so is more likely to be accurate.

    o   Printing of arrays now honours the 'right' argument if there
    are more than two dimensions.

    o   Tabular printing of numbers now has headers right-justified, as
    they were prior to version 1.7.0 (spotted by Rob Baer).

    o   Lazy-loading databases are now cached in memory at first use:
    this enables R to run much faster from slow file systems such
    as USB flash drives.  There is a small (less than 2Mb)
    increase in default memory usage.

    o   The implicit class structure for numeric vectors has been
    changed, so that integer/real vectors try first methods for
    class "integer"/"double" and then those for class "numeric".

    The implicit classes for matrices and arrays have been changed
    to be "matrix"/"array" followed by the class(es) of the
    underlying vector.

    o   splines::splineDesign() now allows the evaluation of a B-spline
    basis everywhere instead of just inside the "inner" knots, by
    setting the new argument `outer.ok = TRUE'.

    o   Hashing has been tweaked to use half as much memory as before.

    o   Readline is not used for tilde expansion when R is run with
    --no-readline, nor from embedded applications.  Then "~name"
    is no longer expanded, but "~" still is.

    o   The regular expression code is now based on that in glibc 2.3.3.
    It has stricter conformance to POSIX, so metachars such as
    { } + * may need to be escaped where before they did not
    (but could have been).

    o   New encoding 'TeXtext.enc' improves the way postscript() works
    with Computer Modern fonts.

    o   Replacement in a non-existent column of a data frame tries
    harder to create a column of the correct length and so avoid a
    corrupt data frame.

    o   For Windows and readline-based history, the saved file size is
    re-read from R_HISTSIZE immediately before saving.

    o   Collected warnings during start-up are now printed before the
    initial prompt rather than after the first command.

    o   Changes to package 'grid':

    - preDrawDetails(), drawDetails(), and postDrawDetails()
      methods are now recorded on the graphics engine
      display list.   This means that calculations within these
      methods are now run when a device is resized or
      when output is copied from one device to another.

    - Fixed bug in grid.text() when 'rot' argument has length 0.
      (privately reported by Emmanuel Paradis)

    - New getNames() function to return just the names of all top-level
      grobs on the display list.

    - Recording on the grid display list is turned off within
      preDrawDetails(), drawDetails(), and postDrawDetails() methods.

    - Grid should recover better from errors or user-interrupts
      during drawing (i.e., not leave you in a strange viewport
      or with strange graphical parameter settings).

    - New function grid.refresh() to redraw the grid display list.

    - New function grid.record() to capture calculations
      with grid graphics output.

    - grobWidth and grobHeight ("grobwidth" and "grobheight" units)
      for primitives (text, rects, etc, ...) are now
      calculated based on a bounding box for the relevant grob.

      NOTE: this has changed the calculation of the size of a scalar
      rect (or circle or lines).

    - New arguments 'warn' and 'wrap' for function grid.grab()

    - New function grid.grabExpr() which captures the output from
      an expression (i.e., not from the current scene) without
      doing any drawing (i.e., no impact on the current scene).

    - upViewport() now (invisibly) returns the path that it goes up
      (suggested by Ross Ihaka).

    - The 'gamma' gpar has been deprecated (this is a device property
      not a property of graphical objects;  suggested by Ross Ihaka).

    - New 'lex' gpar;  a line width multiplier.

    - grid.text() now handles any language object as mathematical
      annotation (instead of just expressions).

    - plotViewport() has default value for 'margins' argument (that match
      the default value for par(mar)).

    - The 'extension' argument to dataViewport() can now be vector,
      in which case the first value is used to extend the xscale and
      the second value is used to extend the y scale.
      (suggested by Ross Ihaka).

    - All 'just' arguments (for viewports, layouts, rectangles, text)
      can now be numeric values (typically between 0 [left] and 1 [right])
      as well as character values ("left", "right", ...).

      For rectangles and text, there are additional 'hjust' and 'vjust'
      arguments which allow numeric vectors of justification in
      each direction (e.g., so that several pieces of text can have
      different justifications).
      (suggested by Ross Ihaka)

    - New 'edits' argument for grid.xaxis() and grid.yaxis() to
      allow specification of on-the-fly edits to axis children.

    - applyEdit(x, edit) returns x if target of edit (i.e., child
      specified by a gPath) cannot be found.

    - Fix for calculation of length of max/min/sum unit.  Length is
      now (correctly) reported as 1 (was reported as length of first arg).

    - Viewport names can now be any string (they used to have to be a
      valid R symbol).

    - The 'label' argument for grid.xaxis() and grid.yaxis() can now
      also be a language object or string vector, in which case it
      specifies custom labels for the tick marks.


INTERNATIONALIZATION

    o   Unix-alike versions of R can now be used in UTF-8 and other
    multi-byte locales on suitably equipped OSes if configured
    with option --enable-mbcs (which is the default).  [The
    changes to font handling in the X11 module are based on the
    Japanization patches of Ei-ji Nakama.]

    Windows versions of R can be used in `East Asian' locales on
    suitable versions of Windows.

    See the 'Internationalization' chapter in the 'Installation
    and Administration' manual.

    o   New command-line flag --encoding to specify the encoding to
    be assumed for stdin (but not for a console).

    o   New function iconv() to convert character vectors between
    encodings, on those OSes which support this.  See the new
    capabilities("iconv").

    o   The meaning of 'encoding' for a connection has changed: it now
    allows any charset encoding supported by iconv on the
    platform, and can re-encode output as well as input.

    As the new specification is a character string and the old was
    numeric, this should not cause incorrect operation.

    o   New function localeToCharset() to find/guess encoding(s) from
    the locale name.

    o   nchar() returns the true number of bytes stored (including any
    embedded nuls), this being 2 for missing values.  It has an
    optional argument 'type' with possible non-default values
    "chars" and "width" to give the number of characters or the
    display width in columns.

    o   Characters can be entered in hexadecimal as e.g. \x9c, and in
    UTF-8 and other multibyte locales as \uxxxx, \u{xxxx},
    \Uxxxxxxxx or \U{xxxxxxxx}.  Non-printable Unicode characters
    are displayed C-style as \uxxxx or \Uxxxxxxxx.

    o   LC_MONETARY is set to the locale, which affects the result of
    Sys.localeconv(), but nothing else in R itself.  (It could
    affect add-on packages.)

    o   source() now has an 'encoding' argument which can be used to
    make it try out various possible encodings.  This is made use
    of by example() which will convert (non-UTF-8) Latin-1 example
    files in a UTF-8 locale.

    o   read/writeChar() work in units of characters, not bytes.

    o   .C() now accepts an ENCODING= argument where re-encoding is
    supported by the OS.  See `Writing R Extensions'.

    o   delimMatch (tools) now reports match positions and lengths in
    units of characters, not bytes.  The delimiters can be
    strings, not just single ASCII characters.

    o   .Rd files can indicate via a \encoding{} argument the encoding
    that should be assumed for non-ASCII characters they contain.

    o   Phrases in .Rd files can be marked by \enc{}{} to show a
    transliteration to ASCII for use in e.g. text help.

    o   The use of 'pch' in points() now allows for multi-byte character
    sets: in such a locale a glyph can either be specified as a
    multi-byte single character or as a number, the Unicode point.

    o   New function l10n_info() reports on aspects of the
    locale/charset currently in use.

    o   scan() is now aware of double-byte locales such as Shift-JIS
    in which ASCII characters can occur as the second ('trail')
    byte.

    o   Functions sQuote() and dQuote() use the Unicode directional
    quotes if in a UTF-8 locale.

    o   The infrastructure is now in place for C-level error and warning
    messages to be translated and used on systems with Native
    Language Support.  This has been used for the startup message
    in English and to translate Americanisms such as 'color' into
    English: translations to several other languages are under
    way, and some are included in this release.

    See 'Writing R Extensions' for how to make use of this in a
    package: all the standard packages have been set up to do
    translation, and the 'language' 'en@quot' is implemented to
    allow Unicode directional quotes in a UTF-8 locale.

    o   R-level stop(), warning() and message() messages can be
    translated, as can other messages via the new function
    gettext(). Tools xgettext() and xgettext2pot() are provided in
    package tools to help manage error messages.

    gettextf() is a new wrapper to call sprintf() using
    gettext() on the format string.

    o   Function ngettext() allows the management of singular and
    plural forms of messages.


UTILITIES

    o   New functions mirror2html() and checkCRAN().

    o   R CMD check has a new option '--use-valgrind'.

    o   R CMD check now checks that Fortran and C++ files have LF
    line endings, as well as C files.  It also checks Makevars[.in]
    files for portable compilation flags.

    o   R CMD check will now work on a source tarball and prints out
    information about the version of R and the package.

    o   tools:::.install_package_code_files() (used to collate R files
    when installing packages) ensures files are separated by a
    line feed.

    o   vignette() now returns an object of class "vignette" whose
    print() method opens the corresponding PDF file.  The edit()
    method can be used to open the code of the vignette in an
    editor.

    o   R CMD INSTALL on Unix has a new option '--build' matching
    that on Windows, to package as tarball the installed package.

    o   R CMD INSTALL on Unix can now install binary bundles.

    o   R CMD build now changes src files to LF line endings if necessary.

    o   R CMD build now behaves consistently between source and binary
    builds: in each case it prepares a source directory and then
    either packages that directory as a tarball or calls R CMD
    INSTALL --build on the prepared sources.

    This means that R CMD build --binary now respects
    .Rbuildignore and will rebuild vignettes (unless the option
    --no-vignettes is used).  For the latter, it now installs the
    current sources into a temporary library and uses that version
    of the package/bundle to rebuild the vignettes.

    o   R CMD build now reports empty directories in the source tree.

    o   New function write_PACKAGES() in package 'tools' to help with
    preparing local package repositories.  (Based on a contribution
    by Uwe Ligges.)  How to prepare such repositories is
    documented in the 'R Installation and Administration' manual.

    o   package.skeleton() adds a bit more to DESCRIPTION.

    o   Sweave changes:

    - \usepackage[nogin]{Sweave} in the header of an Sweave file
      supresses auto-setting of the graphical parameter like width
      of graphics.

    - The new \SweaveInput{} command works similar to LaTeX's
      \input{} command.

    - Option value strip.white=all strips all blank lines from the
      output of a code chunk.

    - Code chunks with eval=false are commented out by Stangle() and
      hence no longer tested by R CMD check.


DOCUMENTATION

    o   File doc/html/faq.html no longer exists, and doc/manual/R-FAQ.html
    (which has active links to other manuals) is used instead.
    (If makeinfo >= 4.7 is not available, the version on CRAN is
    linked to.)

    o   Manual 'Writing R Extensions' has further details on writing
    new front-ends for R using the new public header files.

    o   There are no longer any restrictions on characters in the
    \name{} field of a .Rd file: in particular _ is supported.


C-LEVEL FACILITIES

    o   There are new public C/C++ header files Rinterface.h and
    R_ext/RStartup.h for use with external GUIs.

    o   Added an onExit() function to graphics devices, to be executed
    upon user break if non-NULL.

    o   ISNAN now works even in C++ code that undefines the 'isnan' macro.

    o   R_alloc's limit on 64-bit systems has been raised from just
    under 2^31 bytes (2Gb) to just under 2^34 (16Gb), and is now checked.

    o   New math utility functions  log1pmx(x), lgamma1p(x),
    logspace_add(logx, logy), and logspace_sub(logx, logy).


DEPRECATED & DEFUNCT

    o   The aqua module for MacOS X has been removed: --with-aqua now
    refers to the unbundled Cocoa GUI.

    o   Capabilities "bzip2", "GNOME, "libz" and "PCRE" are defunct.

    o   The undocumented use of UseMethod() with no argument was
    deprecated in 2.0.1 and is now regarded as an error.

    o   Capability "IEEE754" is deprecated.

    o   The 'CRAN' argument to update.packages(), old.packages(),
    new.packages(), download.packages() and install.packages() is
    deprecated in favour of 'repos', which replaces it as a
    positional argument (so this is only relevant for calls with
    named args).

    o   The S3 methods for getting and setting names of "dist" objects
    have been removed (as they provided names with a different
    length from the "dist" object itself).

    o   Option "repositories" is no longer used and so not set.

    o   loadURL() is deprecated in favour of load(url()).

    o   delay() is deprecated.  Use delayedAssign() instead.


INSTALLATION CHANGES

    o   New configure option --enable-mbcs to enable support for
    UTF-8 locales, on by default.

    o   R_XTRA_[CF]FLAGS are now used during the configuration tests,
    and [CF]PICFLAGS if --enable-R-shlib was specified.  This
    ensures that features such as inlining are only used if the
    compilation flags specified support them.  (PR#7257)

    o   Files FAQ, RESOURCES, doc/html/resources.html are no longer in
    the SVN sources but are made by 'make dist'.

    o   The GNOME GUI is unbundled, now provided as a package on CRAN.

    o   Configuring without having the recommended packages is now an
    error unless --with-recommended-packages=no (or equivalent) is used.

    o   Configuring without having the X11 headers and libraries is now
    an error unless --with-x=no (or equivalent) is used.

    o   Configure tries harder to find a minimal set of FLIBS.  Under
    some circumstances this may remove from R_LD_LIBRARY_PATH
    path elements that ought to have specified in LDFLAGS (but
    were not).

    o   The C code for most of the graphics device drivers and their
    afm files are now in package grDevices.

    o   R is now linked against ncurses/termlib/termcap only if
    readline is specified (now the default) and that requires it.

    o   Makeinfo 4.7 or later is now required for building the HTML and
    Info versions of the manuals.


PACKAGE INSTALLATION CHANGES

    o   There are new types of packages, identified by the Type field
    in the DESCRIPTION file.  For example the GNOME console is now
    a separate package (on CRAN), and translations can be
    distributed as packages.

    o   There is now support of installing from within R both source and
    binary packages on MacOS X and Windows.  Most of the R
    functions now have a 'type' argument defaulting to
    getOption("pkgType") and with possible values "source",
    "win.binary" and "mac.binary".  The default is "source" except
    under Windows and the CRAN GUI build for MacOS X.

    o   install.packages() and friends now accept a vector of URLs for
    'repos' or 'contriburl' and get the newest available version of
    a package from the first repository on the list in which it is
    found.  The argument 'CRAN' is still accepted, but deprecated.

    install.packages() on Unix can now install from local .tar.gz
    files via repos = NULL (as has long been done on Windows).

    install.packages() no longer asks if downloaded packages
    should be deleted: they will be deleted at the end of the
    session anyway (and can be deleted by the user at any time).

    If the repository provides the information, install.packages()
    will now accept the name of a package in a bundle.

    If 'pkgs' is omitted install.packages() will use a listbox to
    display the available packages, on suitable systems.

    'dependencies' can be a character vector to allow only some
    levels of dependencies (e.g. not "Suggests") to be requested.

    o   There is a new possible value update.packages(ask="graphics")
    that uses a widget to (de)select packages, on suitable systems.

    o   The option used is now getOption("repos") not getOption("CRAN")
    and it is initially set to a dummy value.  Its value can be a
    character vector (preferably named) giving one or several
    repositories.

    A new function chooseCRANmirror() will select a CRAN mirror.
    This is called automatically if the contrib.url() encounters
    the initial dummy value of getOption("repos")

    A new function setRepositories() can be used to create
    getOption("repos") from a (platform-specific) list of known
    repositories.

    o   New function new.packages() to report uninstalled packages
    available at the requested repositories.  This also reports
    incomplete bundles.  It will optionally install new packages.

    o   New function available.packages(), similar to CRAN.packages()
    but for use with multiple repositories.  Both now only report
    packages whose R version requirements are met.

    o   update.packages() and old.packages() have a new option
    'checkBuilt' to allow packages installed under earlier
    versions of R to be updated.

    o   remove.packages() can now remove bundles.

    o   The Contains: field of the DESCRIPTION file of package bundles
    is now installed, so later checks can find out if the bundle
    is complete.

    o   packageStatus() is now built on top of *.packages, and gains a
    'method' argument.  It defaults to the same repositories as
    the other tools, those specified by getOption("repos").


BUG FIXES

    o   Configuring for Tcl/Tk makes use of ${TK_LIB_SPEC} ${TK_LIBS}
    not ${TK_LIB_SPEC} ${TK_XLIBSW}, which is correct for
    recent versions of Tk, but conceivably not for old
    tkConfig.sh files.

    o   detach() was not recomputing the S4 methods for primitives
    correctly.

    o   Methods package now has class "expression" partly fixed in basic
    classes, so S4 classes can extend these (but "expression" is
    pretty broken as a vector class in R).

    o   Collected warnings had messages with unneeded trailing space.

    o   S4 methods for primitive functions must be exported from
    name spaces; this is now done automatically.
    Note that is.primitive() is now in "base", not "methods".

    o   Package grid:

    - Fixed bug in grid.text() when "rot" argument has length 0.
      (reported by Emmanuel Paradis)

    o   .install_package_vignette_index() created an index even in an
    empty 'doc' directory.

    o   The print() method for factors now escapes characters in the
    levels in the same way as they are printed.

    o   str() removed any class from environment objects.

    str() no longer interprets control characters in character
    strings and factor levels; also no longer truncates factor
    levels unless they are longer than 'nchar.max'.
    Truncation of such long strings is now indicated ''outside''
    the string.

    str(<S4.object>) was misleading for the case of a single slot.

    str() now also properly displays S4 class definitions (such as
    returned by getClass().

    o   print.factor(quote=TRUE) was not quoting levels, causing
    ambiguity when the levels contained spaces or quotes.

    o   R CMD check was confused by a trailing / on a package name.

    o   write.table() was writing incorrect column names if the data
    frame contained any matrix-like columns.

    o   write.table() was not quoting row names for a 0-column x.

    o   t(x)'s default method now also preserves names(dimnames(x)) for
    1D arrays 'x'.

    o   r <- a %*% b no longer produces names(dimnames(r)) == c("", "")
    unless one of a or b has named dimnames.

    o   Some .Internal functions that were supposed to return invisibly
    did not. This was behind PR#7397 and PR#7466.

    o   eval(expr, NULL, encl) now looks up variables in encl, as
    eval(expr, list(), encl) always did

    o   Coercing as.data.frame(NULL) to a pairlist caused an error.

    o   p.adjust(p, ..) now correctly works when `p' contains NAs (or when
    it is of length 0 or length 2 for method = "hommel").

    o   'methods' initialization was calling a function intended for
    .Call() with .C().

    o   optim() needed a check that the objective function returns a
    value of length 1 (spotted by Ben Bolker).

    o   X11() was only scaling its fonts to pointsize if the dpi
    was within 0.5 of 100dpi.

    o   X11() font selection was looking for any symbol font, and
    sometimes got e.g. bold italic if the server has such a font.

    o   dpois(*, lambda=Inf) now returns 0 (or -Inf for log).

    o   Using pch="" gave a square (pch=0)!  Now it is regarded as the
    same as NA, which was also undocumented but omits the point.

    o   Base graphics now notices (ab)lines which have a zero
    coordinate on log scale, and omits them.  (PR#7559)

    o   stop() and warning() now accept NULL as they are documented
    to do (although this seems of little use and is equivalent to "").

    o   weighted.mean() now checks the length of the weight vector w.

    o   getAnywhere() was confused by names with leading or trailing dots
    (spotted by Robert McGehee)

    o   eval() was not handling values from return() correctly.

    o   par(omd) is now of the form c(x1, x2, y1, y2) to match the
    documentation and for S-PLUS compatibility.

    [Previously, par(omd) was of the form c(bottom, left, top, right)
     like par(oma) and par(omi)]

    o   formatC() did not check its 'flag' argument, and could
    segfault if it was incorrect. (PR#7686)

    o   Contrasts needed to be coerced to numeric (e.g. from integer)
    inside model.matrix.  (PR#7695)

    o   socketSelect() did not check for buffered input.

    o   Reads on a non-blocking socket with no available data were
    not handled properly and could result in a segfault.

    o   The "aovlist" method for se.contrast() failed in some very
    simple cases that were effectively not multistratum designs,
    e.g. only one treatment occurring in only one stratum.

    o   pgamma() uses completely re-written algorithms, and should work
    for all (even very extreme) arguments; this is based on Morten
    Welinder's contribution related to PR#7307.

    o   dpois(10, 2e-308, log=TRUE) and similar cases gave -Inf.

    o   x <- 2^(0:1000);    plot(x, x^.9, type="l", log="xy")# and
    x <- 2^-(1070:170); plot(x, x^.9, type="l", log="xy")# now both work

    o   summary.lm() asked for a report on a reasonable occurrence, but
    the check failed to take account of NAs.

    o   lm() was miscalculating 'df.residual' for empty models with a
    matrix response.

    o   summary.lm() now behaves more sensibly for empty models.

    o   plot.window() was using the wrong sign when adjusting
    xlim/ylim for positive 'asp' and a reversed axis.

    o   If malloc() fails when allocating a large object the allocator now
    does a gc and tries the malloc() again.

    o   packageSlot() and getGroupMembers() are now exported from the
    'methods' package as they should from documentation and the
    Green Book.

    o   rhyper() was giving numbers slightly too small, due to a bug in the
    original algorithm.  (PR#7314)

    o   gsub() was sometimes incorrectly matching ^ inside a string,
    e.g.  gsub("^12", "x", "1212") was "xx".

    o   [g]sub(perl = TRUE) was giving random results for a 0-length
    initial match.  (PR#7742)

    o   [g]sub was ignoring most 0-length matches, including all initial
    ones.  Note that substitutions such as gsub("[[:space:]]*", "
    ", ...) now work as they do in 'sed' (whereas the effect was
    previously the same as gsub("[[:space:]]+", " ", ...)).
    (In part PR#7742)

    o   Promises are now evaluated when extracted from an environment
    using '$' or '[[ ]]'.

    o   reshape(direction="wide") had some sorting problems when
    guessing time points (PR#7669)

    o   par() set 'xaxp' before 'xlog' and 'yaxp' before 'ylog',
    causing PR#831.

    o   The logic in tclRequire() to check the availability of a Tcl
    package turned out to be fallible.  It now uses a try()-and-see
    mechanism instead.

    o   Opening a unz() connection on a non-existent file left a file
    handle in use.

    o   "dist" objects of length 0 failed to print.

    o   INSTALL and the libR try harder to find a temporary directory
    (since there might be one left over with the same PID).

    o   acf() could cause a segfault with some datasets.  (PR#7771)

    o   tan(1+LARGEi) now gives 0+1i rather than 0+NaNi (PR#7781)

    o   summary(data.frame(mat = I(matrix(1:8, 4)))) does not go into
    infinite recursion anymore.

    o   writeBin() performed byte-swapping incorrectly on complex
    vectors, also swapping real and imaginary parts. (PR#7778)

    o   read.table() sometimes discarded as blank lines containing
    only white space, even if sep=",".



    **************************************************
    *                        *
    *          2.0 SERIES NEWS           *
    *                        *
    **************************************************


        CHANGES IN R VERSION 2.0.1 patched



NEW FEATURES

    o   warnings() now looks only in the workspace for `last.warning'
    (suggested by PR#7363).

    o   The search for browsers now starts with firefox, and has mozilla
    ahead of netscape.


PACKAGE INSTALLATION CHANGES

    o   A package DESCRIPTION file which contains a Built field (it
    should not!) is now worked around, with loud warnings.


BUG FIXES

    o   split() was accepting raw and list vectors as input, but
    not populating the output correctly.

    split() now handles vectors with names internally and so is
    almost as fast as on vectors without names (and maybe 100x
    faster than before).

    o   subset() now throws an error if its 'subset' argument is not
    logical whereas it could appear to work and give wrong answers
    with e.g. a numeric 'subset' argument.

    o   sum(), max(), min(), prod(), any() and all() were incorrectly
    using partial matching for their na.rm argument.

    o   kmeans() now ensures that the initial cluster centers it chooses
    are distinct.

    o   text()'s default method could segfault if passed 0-length coordinates.

    o   R CMD INSTALL (Unix) was sometimes leaving a temporary dir
    behind.  (PR#7230)

    o   read.table() could fail when the row.names were looking like numbers.

    o   seq(length= <n>) now always returns "integer" storage.

    o   labels.lm was broken. (PR#7417)

    o   The hashing used for character vectors in object.size() was
    inefficient for vectors with thousands of identical values not
    sharing the same storage (an unusual case).

    o   data.matrix() now warns if applied to a data frame with
    classed columns.

    o   Plotting histograms where the expression used to deparse to
    multiple lines now gives a sensible default title.  (PR#7421)

    o   The print method for getDLLRegisteredRoutines() was badly
    designed, and failed for all the standard packages' DLLs.

    o   getDLLRegisteredRoutines.character() was broken.

    o   write.table(x, row.names=FALSE) was incorrect for a 0-column x.

    o   write.table() assumed dec = "," if it was not ".".

    o   morley.tab used by R-intro was not installed on Unix since 2.0.0
    (it was on Windows).

    o   gsub(perl=TRUE) returned a string which printed with trailing
    garbage if there was a match at the beginning whose
    replacement was shorter.  (PR#7479)

    Similarly, the result was truncated if a replacement at the
    beginning was longer.

    o   Some PDF readers do not define PDFDocEncoding, so pdf()'s
    ISOLatin1 encoding is now derived from WinAnsi rather than
    PDFDocEncoding.

    o   xfig() had not been updated for the 2.0.0 alpha changes.
    Using more than one plot with a non-white background and with
    onefile=FALSE could segfault.

    o   sprintf() did not check for buffer overflow on character
    strings.  (PR#7554)

    o   The error message when evaluation depth was exceeded itself
    caused an error in deparsing and so was not shown.

    o   The 'datalist' file was not being used during installation
    (2.0.1 only).

    o   replications() was wrongly reporting lack of balance in
    designs with interactions (and had since unique() gained a
    matrix method).



        CHANGES IN R VERSION 2.0.1


NEW FEATURES

    o   Platform equivalence in library() is tested by a new function
    testPlatformEquivalence() which ignores the 'vendor' field and
    can be customized by cognescenti.

    o   The assignment form of split() allows recycling of vectors
    within the value list. In particular, things like
        split(x, g) <- lapply(split(x, g), mean)
    now work


DOCUMENTATION

    o   Manual `Writing R Extensions' has new sections on writing
    portable packages and on writing new front-ends for R -- the
    latter will be more comprehensive in R 2.1.0 which has new
    public header files.


DEPRECATED & DEFUNCT

    o   The aqua module in MacOS X is deprecated.

    o   Capabilities "bzip2", "GNOME, "libz" and "PCRE" are deprecated.

    o   The GNOME GUI on Unix-alikes is deprecated as part of R;
    it will be available in another form as from R 2.1.0.

    o   The undocumented use of UseMethod() with no argument is now
    formally deprecated.


INSTALLATION CHANGES

    o   Building on Alpha OSF/1 no longer forces the C flag -std1,
    which appears to be no longer needed.  (PR#7257)

    o   The compiler flag -mieee-fp is no longer used on i386 Linux
    (these days it is only passed to the linker and was only
    invoked for compilation steps).

    o   -D__NO_MATH_INLINES is only used on older ix86 glibc-based
    systems which need it (tested at configure time).  This leads to
    small improvements in speed and accuracy on modern systems.

    o   If makeinfo >= 4.5 is not available, warnings are given that
    some of the HTML manuals will be missing, and the index page
    given by help.start() will link to CRAN versions of those manuals.

    o   Files aclocal.m4 and acinclude.m4 used in maintainer builds
    are not longer included in the distribution.


C-LEVEL FACILITIES

    o   It was not clear in 'Writing R Extensions' that some of the
    entry points in the 'Utilities' section were not declared in
    <R.h> (they were in <R_ext/Applic.h>).  Now all the entry
    points in that section are declared in <R_ext/Utils.h>,
    included by <R.h>.


BUG FIXES

    o   The grid.grab() function in package grid would throw an error
    if there were no viewports pushed (now returns NULL).

    o   model.frame.default() takes row names from the response
    variable if that has suitable names and there is no 'data'
    argument.  (This follows S but was not previously implemented
    in R.)

    o   write.table() was not respecting the 'dec' argument for complex
    numbers.

    o   write.table() printed a mixture of numeric and complex numbers
    as all complex.  (PR#7260)

    o   R CMD INSTALL failed with versioned installs on packages which
    save images (only).

    o   dlogis() gave NaN not 0 for large negative arguments.

    o   Importing from another name space was broken for versioned
    installs, incorrectly reporting something like
    "package 'imported_from' does not have a name space".

    o   The GNOME interface under Linux/Unix was broken. (PR#7276)

    o   For the jpeg/png devices under Linux/Unix, under certain rare
    circumstances clipping needed to be cleared before starting a
    new page.  (PR#7270, which has been the case since the devices
    were introduced in 1.1.0.)

    o   First lattice plot (first grid.newpage() call) did not start
    a new page IF there had been a previous traditional graphics
    plot (on the same device).

    o   Using install.packages() to install the same package to more
    than one library gave an incorrect warning message.  (If there
    were two or more such packages it might give an error.)

    o   .packages(all.available=TRUE) returned packages with an invalid
    version field in their DESCRIPTION whereas .find.packages() and
    packageDescription() did not.  Now all do not.

    o   packageDescription() now correctly reports that a package does
    not exist, rather than that its DESCRIPTION file is 'missing
    or broken'.

    o   'make dist' from builddir != sourcedir was copying not linking
    recommended packages to *.tgz.

    o   Slots in prototype objects can inherit from locally defined
    classes (which were not being found correctly before).

    o   Several fixes to the behavior of as() when there are either
    coerce= or replace= methods supplied in a call to
    setIs(). Related fixes to setIs() to handle correctly previous
    methods, if there were any.

    o   splinefun(1[0], 1[0])(1) doesn't segfault anymore (PR#7290).
    spline() and splinefun() now also work with missing values by
    omiting them.

    o   ecdf() was failing on inputs containing NAs. (Part of PR#7292)

    o   tools:::.install_package_description was splitting the Built:
    field across lines on platforms with very long names.

    o   capabilities() was wrong for the Aqua GUI on MacOS X.

    o   Using Rprof() with a non-writable 'file' argument is now a
    non-fatal error and does not abort R.

    o   binom.test() did not deparse its arguments early enough such
    that the reported data were ugly if x was a table.

    o   Systems based on glibc, including those using R's substitute for
    strptime, were handling strptime("2001", "%Y") incorrectly, in
    some cases crashing.  R's substitute code has been corrected
    (but problems may remain if glibc is used).  See the ?strptime
    for what should happen (which is system-specific).

    o   untrace() after trace() failed if package 'methods' was attached.
    (PR#7301)

    o   summary.stepfun() was reporting for n > 6 summaries of the
    knots and levels as the actual values.  Both print() and
    summary() methods called the constant values "step heights",
    although they were not the heights of the steps.

    o   is.na/is.nan() were giving spurious warnings if applied to a
    raw vector.

    o   is.atomic() gave incorrect result (false) for a raw vector.

    o   rank() and order() accepted raw and list inputs, but did not
    give a sensible answer (always 1:n).  Similarly, partial sorts
    of a raw vector were accepted but did nothing.

    o   require() without a version argument tried for an unversioned
    load of a package even though a versioned install was already
    loaded.  This often led to a message that a required package
    was being loaded when it was not actually being loaded.

    o   str(<S4.object>) made use of attributes() instead of slot(),
    and hence didn't properly print NULL slots.

    o   contrib.url() now handles URLs ending in '/' correctly.

    o   str() removed any class from externalptr objects.

    o   logLik() and hence AIC() failed or gave incorrect answers
    on "lm" fits with na.action = na.exclude (and perhaps other
    na.actions's except na.omit and na.fail).

    o   pmax() and pmin() sometimes used NAs in internal subassignments,
    and sometimes these failed.

    o   Subassigning an expression, e.g. expr[2] <- 1, could leave an
    invalid object and so cause a segfault. (PR#7326)

    o   download/install.packages() would misbehave if there was more
    than one version of a package in a repository.

    o   sort(partial=) silently ignored some other arguments: using
    'decreasing' or 'index.return' or supplying a factor are now
    errors.

    o   The ave() function had trouble if the grouping contained
    unused levels.

    o   read.fwf() got confused by skip > 0 and could infinite loop
    under some circumstances.  (PR#7350)

    o   upgrade(x, ask = FALSE) was broken for a "packageStatus" object.

    o   Class "raw" had been omitted from the list of basic classes in
    the "methods" package and so could not be used in S4 classes.

    o   Function getGroupMembers(), part of the definition of S4
    classes, had been promised for release 2.0, but slipped through.

    o   toLatex(sessionInfo()) produced incorrect LaTeX on some
    platforms due to special characters in the platform
    identifier.




    *********************************************************
    *                           *
    *   News of 1.x.y and 2.0.0 is in file `ONEWS'  *
    *   News of 1.0.0 and earlier is in file `OONEWS'   *
    *                           *
    *********************************************************