r-cran-matrixstats 0.52.2-2 / usr / lib / R / site-library / matrixStats / NEWS

Package: matrixStats
====================

Version: 0.52.2 [2017-04-13]

BUG FIX:

 o Several of the row- and column-based functions would core dump R if the
   matrix was of a data type other than logical, integer, or numeric, e.g.
   character or complex.  This is now detected and an informative error is
   produced instead.  Similarly, some vector-based functions could potentially
   core dump R or silently return a nonsense result.  Thank you Hervé Pagès,
   Bioconductor Core, for the report.
 
DEPRECATED AND DEFUNCT:

 o rowVars(..., method = "0.14.2") that was added for very unlikely needs of
   backward compatibility of an invalid degree-of-freedom term is deprecated.
   
 
Version: 0.52.1 [2017-04-04]

BUG FIX:

 o The package test on matrixStats:::benchmark() tried to run even if
   not all suggested packages were available.
 

Version: 0.52.0 [2017-04-03]

SIGNIFICANT CHANGES:

 o Since anyNA() is a built-in function since R (>= 3.1.0), please use that
   instead of anyMissing() part of this package.  The latter will eventually
   be deprecated.  For consistency with the anyNA() name, colAnyNAs() and
   rowAnyNAs() are now also available replacing the identically
   colAnyMissings() and rowAnyMissings() functions, which will also be
   deprecated in a future release.

 o meanOver() was renamed to mean2() and sumOver() was renamed to sum2().

NEW FEATURES:

 o Added colSums2() and rowSums2() which work like colSums() and rowSums()
   of the base package but also supports efficient subsetting via optional
   arguments 'rows' and 'cols'.

 o Added colMeans2() and rowMeans2() which work like colMeans() and rowMeans()
   of the base package but also supports efficient subsetting via optional
   arguments 'rows' and 'cols'.

 o Functions colDiffs() and rowDiffs() gained argument 'dim.'.

 o Functions colWeightedMads() and rowWeightedMads() gained arguments
   'constant' and 'center'.  The current implementation only support scalars
   for these arguments, which means that the same values are applied to all
   columns and rows, respectively.  In previous version a hard-to-understand
   error would be produced if 'center' was of length greater than one; now
   an more informative error message is given.

 o Package is now silent when loaded; it no longer displays a startup message.
 
SOFTWARE QUALITY:

 o Continuous-integration testing is now also done on macOS, in addition to
   Linux and Windows.
   
 o ROBUSTNESS: Package now registers the native API using also
   R_useDynamicSymbols().

CODE REFACTORING:

 o Cleaned up native low-level API and renamed native source code files
   to make it easier to navigate the native API.
 
 o Now using roxygen for help and NAMESPACE (was R.oo::Rdoc).

BUG FIXES:

 o rowAnys(x) on numeric matrices 'x' would return rowAnys(x == 1) and
   not rowAnys(x != 0).  Same for colAnys(), rowAlls(), and colAlls().
   Thanks Richard Cotton for reporting on this.

 o sumOver(x) and meanOver(x) would incorrectly return -Inf or +Inf if the
   intermediate sum would have that value, even if one of the following
   elements would turn the intermediate sum into NaN or NA, e.g. with 'x'
   as c(-Inf, NaN), c(-Inf, +Inf), or c(+Inf, NA).
   
 o WORKAROUND: Benchmark reports generated by matrixStats:::benchmark() would
   use any custom R prompt that is currently set in the R session, which may
   not render very well.  Now it forces the prompt to be the built-in "> " one.

DEPRECATED AND DEFUNCT:

 o The package API is only intended for matrices and vectors of type
   numeric, integer and logical.  However, a few functions would still
   return if called with a data.frame.  This was never intended to work
   and is now an error.  Specifically, functions colAlls(), colAnys(),
   colProds(), colQuantiles(), colIQRs(), colWeightedMeans(),
   colWeightedMedians(), and colCollapse() now produce warnings if called
   with a data.frame.  Same for the corresponding row- functions.
   The use of a data.frame will be produce an error in future releases.

 o meanOver() and sumOver() are deprecated because they were renamed to
   mean2() and sum2(), respectively.

 o Previously deprecated (and ignored) argument 'flavor' of colRanks() and
   rowRanks() is now defunct.


Version: 0.51.0 [2016-10-08]

PERFORMANCE AND MEMORY:

 o SPEEDUP / CLEANUP: rowMedians() and colMedians() are now plain functions.
   They were previously S4 methods (due to a Bioconductor legacy).  The
   package no longer imports the methods package.

 o SPEEDUP: Now native API is formally registered allowing for faster lookup
   of routines from R.
 
 
Version: 0.50.2 [2016-04-24]

BUG FIXES:

 o Package now installs on R (>= 2.12.0) as claimed.  Thanks to Mikko Korpela
   at Aalto University School of Science, Finland, for troubleshooting and
   providing a fix.

 o logSumExp(c(-Inf, -Inf, ...)) would return NaN rather than -Inf. Thanks to
   Jason Xu (University of Washington) for reporting and Brennan Vincent for
   troubleshooting and contributing a fix.

 
Version: 0.50.1 [2015-12-14]

BUG FIXES:

 o The Undefined Behavior Sanitizer (UBsan) reported on a memcall(src, dest, 0)
   call when dest == null.  Thanks to Brian Ripley and the CRAN check tools for
   catching this. We could reproduce this with gcc 5.1.1 but not with gcc 4.9.2.
 
 
Version: 0.50.0 [2015-12-13]

NEW FEATURES:

 o MAJOR FEATURE UPDATE: Subsetting arguments 'idxs', 'rows' and 'cols' were
   added to all functions such that the calculations are performed on the
   requested subset while avoiding creating a subsetted copy, i.e.
   rowVars(x, cols = 4:6) is a much faster and more memory efficient version
   than rowVars(x[, 4:6]) and even yet more efficient than
   apply(x, MARGIN = 1L, FUN = var). These features were added by Dongcan Jiang,
   Peking University, with support from the Google Summer of Code program.
   A great thank you to Dongcan and to Google for making this possible.
 
 
Version: 0.15.0 [2015-10-26]

NEW FEATURES:

 o CONSISTENCY: Now all weight arguments ('w' and 'W') default to NULL, which
   corresponds to uniform weights.

CODE REFACTORING:

 o ROBUSTNESS: Importing 'stats' functions in namespace.

BUG FIXES:

 o weightedVar(x, w) used the wrong bias correction factor resulting in an
   estimate that was tau too large, where
   tau = ((sum(w) - 1) / sum(w)) / ((length(w) - 1) / length(w)).
   Thanks to Wolfgang Abele for reporting and troubleshooting on this.

 o weightedVar(x) with length(x) = 1 returned 0 no NA. Same for weightedSd().

 o weightedMedian(x, w = NA_real_) returned 'x' rather than NA_real_. This
   only happened for length(w) = 1.

 o allocArray(dim) failed for prod(dim) >= .Machine$integer.max.

DEPRECATED AND DEFUNCT:

 o CLEANUP: Defunct argument 'centers' for col-/rowMads(); use 'center'.

 
Version: 0.14.2 [2015-06-23]

BUG FIXES:

 o x_OP_y() and t_tx_OP_y() would return garbage on Solaris SPARC (and possibly
   other architectures as well) when input was integer and had missing values.
 
 
Version: 0.14.1 [2015-06-17]

BUG FIXES:

 o product(x, na.rm = FALSE) for integer 'x' with both zeros and NAs returned
   zero rather than NA.

 o weightedMean(x, w, na.rm = TRUE) did not handle missing values in 'x'
   properly, if it was an integer.  It would also return NaN if there were
   weights 'w' with missing values, whereas stats::weighted.mean() would skip
   such data points.  Now weightedMean() does the same.

 o (col|row)WeightedMedians() did not handle infinite weights as
   weightedMedian() does.

 o x_OP_y(x, y, OP, na.rm = FALSE) returned garbage iff 'x' or 'y' had
   missing values of type integer.

 o rowQuantiles() and rowIQRs() did not work for single-row matrices.
   Analogously for the corresponding column functions.

 o rowCumsums(), rowCumprods() rowCummins(), and rowCummaxs(), accessed
   out-of-bound elements for Nx0 matrices where N > 0.  The corresponding
   column methods has similar memory errors for 0xK matrices where K > 0.

 o anyMissing(list(NULL)) returned NULL; now FALSE.

 o rowCounts() resulted in garbage if a previous column had NAs (because it
   forgot to update index kk in such cases).

 o rowCumprods(x) handled missing values and zeros incorrectly for integer
   'x (not double); a zero would trump an existing missing value causing the
   following cumulative products to become zero.  It was only a zero that
   trumped NAs; any other integer would work as expected.  Note, this bug
   was not in colCumprods().

 o rowAnys(x, value, na.rm = FALSE) did not handle missing values in a numeric
   'x' properly.  Similarly, for non-numeric and non-logical 'x', row- and
   colAnys(), row- and colAlls(), anyValue() and allValue() did not handle
   when 'value' was a missing value.

 o All of the above bugs were identified and fixed by Dongcan Jiang (Peking
   University, China), who also added corresponding unit tests.
 
 
Version: 0.14.0 [2015-02-13]

SIGNIFICANT CHANGES:

 o CLEANUP: anyMissing() is no longer an S4 generic.  This was done as part of
   the migration of making all functions of matrixStats plain R functions,
   which minimizes calling overhead and it will also allow us to drop 'methods'
   from the package dependencies.  I've scanned all CRAN and Bioconductor
   packages depending on matrixStats and none of them relied on anyMissing()
   dispatching on class, so hopefully this move has little impact. The only
   remaining S4 methods are now colMedians() and rowMedians().

NEW FEATURES:

 o CONSISTENCY: Renamed argument 'centers' of col- and rowMads() to 'center'.
   This is consistent with col- and rowVars().

 o CONSISTENCY: col- and rowVars() now use na.rm = FALSE as the default
   (na.rm = TRUE was mistakenly introduced as the default in v0.9.7).

PERFORMANCE AND MEMORY:

 o SPEEDUP: The check for user interrupts at the C level is now done less
   frequently of the functions.  It does every k:th iteration, where
   k = 2^20, which is tested for using (iter % k == 0).  It turns out, at
   least with the default compiler optimization settings that I use, that
   this test is 3 times faster if k = 2^n where n is an integer.  The
   following functions checks for user interrupts: logSumExp(),
   (col|row)LogSumExps(), (col|row)Medians(),, (col|row)Mads(),
   (col|row)Vars(), and (col|row)Cum(Min|Max|prod|sum)s().

 o SPEEDUP: logSumExp(x) is now faster if 'x' does not contain any missing
   values.  It is also faster if all values are missing or the maximum value
   is +Inf - in both cases it can skip the actual summation step.

SOFTWARE QUALITY:

 o ROBUSTNESS/TESTS: Package tests cover 96% of the code (was 91%).

CODE REFACTORING:

 o CLEANUP: Package no longer depends on R.methodsS3.

BUG FIXES:

 o all() and any() flavored methods on non-numeric and non-logical (e.g.
   character) vectors and matrices with na.rm = FALSE did not give results
   consistent with all() and any() if there were missing values.  For
   example, with x <- c("a", NA, "b") we have all(x == "a") == FALSE and
   any(x == "a") == TRUE whereas our corresponding methods would return NA in
   those cases.  The methods fixed are allValue(), anyValue(), col- and
   rowAlls(), and col- and rowAnys().  Added more package tests to cover
   these cases.

 o logSumExp(x, na.rm = TRUE) would return NA if all values were NA and
   length(x) > 1.  Now it returns -Inf for all length(x):s.
 
 
Version: 0.13.1 [2015-01-21]

BUG FIXES:

 o diff2() with differences >= 3 would *read* spurious values beyond the
   allocated memory.  This error, introduced in 0.13.0, was harmless in the
   sense that the returned value was unaffected and still correct.  Thanks
   to Brian Ripley and the CRAN check tools for catching this.  I could
   reproduce it locally with 'valgrind'.
 
 
Version: 0.13.0 [2015-01-20]

SIGNIFICANT CHANGES:

 o SPEEDUP/CLEANUP: Turned several S3 and S4 methods into plain R functions,
   which decreases the overhead of calling the functions.  After this there
   are no longer any S3 methods.  Remaining S4 methods are anyMissing() and
   rowMedians().

NEW FEATURES:

 o Added weightedMean(), which is ~10 times faster than stats::weighted.mean().

 o Added count(x, value) which is a notably faster than sum(x == value).  This
   can also be used to count missing values etc.

 o Added allValue() and anyValue() for all(x == value) and any(x == value).

 o Added diff2(), which is notably faster than base::diff() for vectors, which
   it is designed for.

 o Added iqrDiff() and (col|row)IqrDiffs().

 o CONSISTENCY: Now rowQuantiles(x, na.rm = TRUE) returns all NAs for rows
   with missing values.  Analogously for colQuantiles(), colIQRs(), rowIQRs()
   and iqr().  Previously, all these functions gave an error saying missing
   values are not allowed.

 o COMPLETENESS: Added corresponding "missing" vector functions for already
   existing column and row functions.  Similarly, added "missing" column and
   row functions for already existing vector functions, e.g. added iqr() and
   count() to complement already existing (col|row)IQRs() and (col|row)Counts()
   functions.

 o ROBUSTNESS: Now column and row methods give slightly more informative error
   messages if a data.frame is passed instead of a matrix.

DOCUMENTATION:

 o DOCUMENTATION: Added vignette summarizing available functions.

PERFORMANCE AND MEMORY:

 o SPEEDUP: (col|row)Diffs() are now implemented in native code and notably
   faster than diff() for matrices.

 o SPEEDUP: Made binCounts() and binMeans() a bit faster.

 o SPEEDUP: Implemented weightedMedian() in native code, which made it ~3-10
   times faster.   Dropped support for ties = "both", because it would have to
   return two values in case of ties, which made the API unnecessarily
   complicated.  If really needed, then call the function twice with
   ties = "min" and ties = "max".

 o SPEEDUP: (col|row)Anys() and (col|row)Alls() is now notably faster compared
   to previous versions.

CODE REFACTORING:

 o CLEANUP: In the effort of migrating anyMissing() into a plain R function,
   the specific anyMissing() implementations for data.frame:s and and list:s
   were dropped and is now handled by anyMissing() for "ANY", which is the only
   S4 method remaining now.  In a near future release, this remaining "ANY"
   method will turned into a plain R function and the current S4 generic will
   be dropped.  We know of know CRAN and Bioconductor packages that relies on
   it being a generic function.  Note also that since R (>= 3.1.0) there is a
   base::anyNA() function that does the exact same thing making anyMissing()
   obsolete.

BUG FIXES:

 o weightedMedian(..., ties = "both") would give an error if there was a tie.
   Added package test for this case.
 
 
Version: 0.12.2 [2014-12-07]

BUG FIXES:

 o CODE FIX: The native code for product() on integer vector incorrectly used
   C-level abs() on intermediate values despite those being doubles requiring
   fabs().  Despite this, the calculated product would still be correct (at
   least when validated on several local setups as well as on the CRAN servers).
   Again, thanks to Brian Ripley for pointing out another invalid integer-double
   coersion at the C level.
 
 
Version: 0.12.1 [2014-12-06]

SOFTWARE QUALITY:

 o ROBUSTNESS: Updated package tests to check methods in more scenarios,
   especially with both integer and numeric input data.

BUG FIXES:

 o (col|row)Cumsums(x) where 'x' is integer would return garbage for columns
   (rows) containing missing values.

 o rowMads(x) where 'x' is numeric (not integer) would give incorrect results
   for rows that had an *odd* number of values (no ties).  Analogously issues
   with colMads(). Added package tests for such cases too.  Thanks to Brian
   Ripley and the CRAN check tools for (yet again) catching another coding
   mistake.  Details: This was because the C-level calculation of the absolute
   value of residuals toward the median would use integer-based abs() rather
   than double-based fabs(). Now it fabs() is used when the values are double
   and abs() when they are integers.
 
 
Version: 0.12.0 [2014-12-05]

 o Submitted to CRAN.
 
 
Version: 0.11.9 [2014-11-26]

NEW FEATURES:

 o Added (col|row)Cumsums(), (col|row)Cumprods(), (col|row)Cummins(), and
   (col|row)Cummaxs().

BUG FIXES:

 o (col|row)WeightedMeans() with all zero weights gave mean estimates with
   values 0 instead of NaN.
 
 
Version: 0.11.8 [2014-11-25]

PERFORMANCE AND MEMORY:

 o SPEEDUP: Implemented (col|row)Mads(), (col|row)Sds() and (col|row)Vars() in
   native code.

 o SPEEDUP: Made (col|row)Quantiles(x) faster for 'x' without missing values
   (and default type = 7L quantiles).  It should still be implemented in
   native code though.

 o SPEEDUP: Made rowWeightedMeans() faster.

BUG FIXES:

 o (col|row)Medians(x) when 'x' is integer would give invalid median values in
   case (a) it was calculated as the mean of two values ("ties"), and (b) the
   sum of those values where greater than .Machine$integer.max.  Now such ties
   are calculated using floating point precision.  Add lots of package tests.
 
 
Version: 0.11.6 [2014-11-16]

PERFORMANCE AND MEMORY:

 o SPEEDUP: Now (col|row)Mins(), (col|row)Maxs() and (col|row)Ranges() are
   implemented in native code providing a significant speedup.

 o SPEEDUP: Now colOrderStats() also is implemented in native code, which
   indirectly makes colMins(), colMaxs() and colRanges() faster.

 o SPEEDUP: colTabulates(x) no longer uses rowTabulates(t(x)).

 o SPEEDUP: colQuantiles(x) no longer uses rowQuantiles(t(x)).

DEPRECATED AND DEFUNCT:

 o CLEANUP: Argument 'flavor' of (col|row)Ranks() is now ignored.

 
Version: 0.11.5 [2014-11-15]

SIGNIFICANT CHANGES:

 o (col|row)Prods() now uses default method = "direct" (was "expSumLog").
 
PERFORMANCE AND MEMORY:

 o SPEEDUP: Now colCollapse(x) no longer utilizes rowCollapse(t(x)).  Added
   package tests for (col|row)Collapse().

 o SPEEDUP: Now colDiffs(x) no longer uses rowDiffs(t(x)). Added package tests
   for (col|row)Diffs().

 o SPEEDUP: Package no longer utilizes match.arg() due to its overhead; methods
   sumOver(), (col|row)Prods() and (col|row)Ranks() were updated.

 
Version: 0.11.4 [2014-11-14]

NEW FEATURES:

 o Added support for vector input to several of the row- and column methods
   as long as the "intended" matrix dimension is specified via argument 'dim'.
   For instance, rowCounts(x, dim = c(nrow, ncol)) is the same as
   rowCounts(matrix(x, nrow, ncol)), but more efficient since it avoids
   creating/allocating a temporary matrix.

PERFORMANCE AND MEMORY:

 o SPEEDUP: Now colCounts() is implemented in native code.  Moreover,
   (col|row)Counts() are now also implemented in native code for logical input
   (previously only for integer and double input).  Added more package tests
   and benchmarks for these functions.
 
 
Version: 0.11.3 [2014-11-11]

SIGNIFICANT CHANGES:

 o Turned sdDiff(),  madDiff(), varDiff(), weightedSd(), weightedVar() and
   weightedMad() into plain functions (were generic functions).

CODE REFACTORING:

 o Removed unnecessary usage of '::'.
 
 
Version: 0.11.2 [2014-11-09]

SIGNIFICANT CHANGES:

 o SPEEDUP: Implemented indexByRow() in native code and it is no longer a
   generic function, but a regular function, which is also faster to call.
   The first argument of indexByRow() has been changed to 'dim' such that one
   should use indexByRow(dim(X)) instead of indexByRow(X) as in the past.
   The latter form is still supported, but deprecated.

NEW FEATURES:

 o Added allocVector(), allocMatrix() and allocArray() for faster allocation
   numeric vectors, matrices and arrays, particularly when filled with
   non-missing values.

DEPRECATED AND DEFUNCT:

 o Calling indexByRow(X) with a matrix 'X' is deprectated.  Instead call it
   with indexByRow(dim(X)).

 
Version: 0.11.1 [2014-11-07]

NEW FEATURES:

 o Better support for long vectors.

 o PRECISION: Using greater floating-point precision in more internal
   intermediate calculations, where possible.

SOFTWARE QUALITY:

 o ROBUSTNESS: Although unlikely, with long vectors support for binCounts()
   and binMeans() it is possible that a bin gets a higher count than what
   can be represented by an R integer (.Machine$integer.max = 2^31-1).  If
   that happens, an informative warning is generated and the bin count is
   set to .Machine$integer.max.  If this happens for binMeans(), the
   corresponding mean is still properly calculated and valid.

CODE REFACTORING:

 o CLEANUP: Cleanup and harmonized the internal C API such there are two
   well defined API levels.  The high-level API is called by R via .Call()
   and takes care of most of the argument validation and construction of
   the return value.  This function dispatch to functions in the low-level
   API based on data type(s) and other arguments.  The low-level API is
   written to work with basic C data types only.

BUG FIXES:

 o Package incorrectly redefined R_xlen_t on R (>= 3.0.0) systems where
   LONG_VECTOR_SUPPORT is not supported.
 
 
Version: 0.11.0 [2014-11-02]

NEW FEATURES:

 o Added sumOver() and meanOver(), which are notably faster versions of
   sum(x[idxs]) and mean(x[idxs]).  Moreover, instead of having to do
   sum(as.numeric(x)) to avoid integer overflow when 'x' is an integer vector,
   one can do sumOver(x, mode = "numeric"), which avoids the extra copy
   created when coercing to numeric (this numeric copy is also twice as large
   as the integer vector).  Added package tests and benchmark reports for
   these functions.
 
 
Version: 0.10.4 [2014-11-01]

PERFORMANCE AND MEMORY:

 o SPEEDUP: Made anyMissing(), logSumExp(), (col|row)Medians(),
   (col|row)Counts() slightly faster by making the native code assign
   the results directly to the native vector instead of to the R vector,
   e.g. ansp[i] = v where ansp = REAL(ans) instead of REAL(ans)[i] = v.

 o Added benchmark reports for anyMissing() and logSumExp().
 
 
Version: 0.10.3 [2014-10-01]

BUG FIXES:

 o binMeans() returned 0.0 instead of NA_real_ for empty bins.
 
 
Version: 0.10.2 [2014-09-01]

BUG FIXES:

 o On some systems, the package failed to build on R (<= 2.15.3) with
   compilation error: "redefinition of typedef 'R_xlen_t'".
 
 
Version: 0.10.1 [2014-06-09]

PERFORMANCE AND MEMORY:

 o Added benchmark reports for also non-matrixStats functions col/rowSums()
   and col/rowMeans().

 o Now all colNnn() and rowNnn() methods are benchmarked in a combined report
   making it possible to also compare colNnn(x) with rowNnn(t(x)).
 
 
Version: 0.10.0 [2014-06-07]

SOFTWARE QUALITY:

 o Relaxed some packages tests such that they assert numerical correctness via
   all.equal() rather than identical().

 o Submitted to CRAN.

BUG FIXES:

 o The package tests for product() incorrectly assumed that the value of
   prod(c(NaN, NA)) is uniquely defined.  However, as documented in
   help("is.nan"), it may be NA or NaN depending on R system/platform.

 
Version: 0.9.7 [2014-06-05]

BUG FIXES:

 o Introduced a bug in v0.9.5 causing col- and rowVars() and hence also
   col- and rowSds() to return garbage.  Add package tests for these now.

 o Submitted to CRAN.
 
 
Version: 0.9.6 [2014-06-04]

NEW FEATURES:

 o Added signTabulate() for tabulating the number of negatives, zeros,
   positives and missing values.  For doubles, the number of negative and
   positive infinite values are also counted.

PERFORMANCE AND MEMORY:

 o SPEEDUP: Now col- and rowProds() utilizes new product() function.

 o SPEEDUP: Added product() for calculating the product of a numeric
   vector via the logarithm.
 
 
Version: 0.9.5 [2014-06-04]

SIGNIFICANT CHANGES:

 o SPEEDUP: Made weightedMedian() a plain function (was an S3 method).

 o CLEANUP: Now only exporting plain functions and generic functions.
 
 o SPEEDUP: Turned more S4 methods into S3 methods, e.g. rowCounts(),
   rowAlls(), rowAnys(), rowTabulates() and rowCollapse().

NEW FEATURES:

 o Added argument 'method' to col- and rowProds() for controlling how the
   product is calculated.

PERFORMANCE AND MEMORY:

 o SPEEDUP: Package is now byte compiled.

 o SPEEDUP: Made rowProds() and rowTabulates() notably faster.

 o SPEEDUP: Now rowCounts(), rowAnys(), rowAlls() and corresponding column
   methods can search for any value in addition to the default TRUE.  The
   search for a matching integer or double value is done in native code,
   which is notably faster (and more memory efficient because it avoids
   creating any new objects).

 o SPEEDUP: Made colVars() and colSds() notably faster and rowVars() and
   rowSds() a slightly bit faster.

 o Added benchmark reports, e.g. matrixStats:::benchmark('colMins').
 
 
Version: 0.9.4 [2014-05-23]

SIGNIFICANT CHANGES:

 o SPEEDUP: Turned several S4 methods into S3 methods, e.g. indexByRow(),
   madDiff(), sdDiff() and varDiff().
 
 
Version: 0.9.3 [2014-04-26]

NEW FEATURES:

 o Added argument 'trim' to madDiff(), sdDiff() and varDiff().
 
 
Version: 0.9.2 [2014-04-04]

BUG FIXES:

 o The native code of binMeans(x, bx) would try to access an out-of-bounds
   value of argument 'y' iff 'x' contained elements that are left of all bins
   in 'bx'.  This bug had no impact on the results and since no assignment was
   done it should also not crash/core dump R.  This was discovered thanks to
   new memtests (ASAN and valgrind) provided by CRAN.
 
 
Version: 0.9.1 [2014-03-31]

BUG FIXES:

 o rowProds() would throw "Error in rowSums(isNeg) : 'x' must be an array of
   at least two dimensions" on matrices where all rows contained at least one
   zero.  Thanks to Roel Verbelen at KU Leuven for the report.
 
 
Version: 0.9.0 [2014-03-26]

NEW FEATURES:

 o Added weighedVar() and weightedSd().
 
 
Version: 0.8.14 [2013-11-23]

PERFORMANCE AND MEMORY:

 o MEMORY: Updated all functions to do a better job of cleaning out temporarily
   allocated objects as soon as possible such that the garbage collector can
   remove them sooner, iff wanted.  This increase the chance for a smaller
   memory footprint.

 o Submitted to CRAN.
 
 
Version: 0.8.13 [2013-10-08]

NEW FEATURES:

 o Added argument 'right' to binCounts() and binMeans() to specify whether
   binning should be done by (u,v] or [u,v).  Added system tests validating
   the correctness of the two cases.

CODE REFACTORING:

 o Bumped up package dependencies.
 
 
Version: 0.8.12 [2013-09-26]

PERFORMANCE AND MEMORY:

 o SPEEDUP: Now utilizing anyMissing() everywhere possible.
 
 
Version: 0.8.11 [2013-09-21]

SOFTWARE QUALITY:

 o ROBUSTNESS: Now importing 'loadMethod' from 'methods' package such that
   'matrixStats' S4-based methods also work when 'methods' is not loaded, e.g.
   when 'Rscript' is used, cf. Section 'Default packages' in
   'R Installation and Administration'.

 o ROBUSTNESS: Updates package system tests such that the can run with only
   the 'base' package loaded.
 
 
Version: 0.8.10 [2013-09-15]

CODE REFACTORING:

 o CLEANUP: Now only importing two functions from the 'methods' package.

 o Bumped up package dependencies.
 
 
Version: 0.8.9 [2013-08-29]

NEW FEATURES:

 o CLEANUP: Now the package startup message acknowledges argument
   'quietly' of library()/require().
 
 
Version: 0.8.8 [2013-07-29]

DOCUMENTATION:

 o The dimension of the return value was swapped in help("rowQuantiles").
 
 
Version: 0.8.7 [2013-07-28]

PERFORMANCE AND MEMORY:

 o SPEEDUP: Made (col|row)Mins() and (col|row)Maxs() much faster.

BUG FIXES:

 o rowRanges(x) on an Nx0 matrix would give an error.  Same for colRanges(x)
   on an 0xN matrix.  Added system tests for these and other special cases.
 
 
Version: 0.8.6 [2013-07-20]

CODE REFACTORING:

 o Bumped up package dependencies.
 
BUG FIXES:

 o Forgot to declare S3 methods (col|row)WeightedMedians().
 
 
Version: 0.8.5 [2013-05-25]

PERFORMANCE AND MEMORY:

 o Minor speedup of (col|row)Tabulates() by replacing rm() calls with NULL
   assignments.
 
 
Version: 0.8.4 [2013-05-20]

DOCUMENTATION:

 o CRAN POLICY: Now all Rd \usage{} lines are at most 90 characters long.
 
 
Version: 0.8.3 [2013-05-10]

PERFORMANCE AND MEMORY:

 o SPEEDUP: binCounts() and binMeans() now uses Hoare's Quicksort for
   presorting 'x' before counting/averaging.  They also no longer test in
   every iteration (== for every data point) whether the last bin has been
   reached or not, but only after completing a bin.
 
 
Version: 0.8.2 [2013-05-02]

DOCUMENTATION:

 o Minor corrections and updates to help pages.
 
 
Version: 0.8.1 [2013-05-02]

BUG FIXES:

 o Native code of logSumExp() used an invalid check for missing value of an
   integer argument.  Detected by Brian Ripley upon CRAN submission.
 
 
Version: 0.8.0 [2013-05-01]

NEW FEATURES:

 o Added logSumExp(lx) and (col|row)LogSumExps(lx) for accurately computing
   of log(sum(exp(lx))) for standalone vectors, and row and column vectors of
   matrices. Thanks to Nakayama (Japan) for the suggestion and contributing a
   draft in R.
 
 
Version: 0.7.1 [2013-04-23]

NEW FEATURES:

 o Added argument 'preserveShape' to colRanks().  For backwardcompatibility
   the default is preserveShape = FALSE, but it may change in the future.

BUG FIXES:

 o Since v0.6.4, (col|row)Ranks() gave the incorrect results for integer
   matrices with missing values.

 o Since v0.6.4, (col|row)Medians() for integers would calculate ties as
   floor(tieAvg).
 
 
Version: 0.7.0 [2013-01-14]

NEW FEATURES:

 o Now (col|row)Ranks() support "max" (default), "min" and "average" for
   argument 'ties.method'.  Added system tests validation these cases.
   Thanks Peter Langfelder (UCLA) for contributing this.
 
 
Version: 0.6.4 [2013-01-13]

NEW FEATURES:

 o Added argument 'ties.method' to rowRanks() and colRanks(), but still only
   support for "max" (as before).

CODE REFACTORING:

 o ROBUSTNESS: Lots of cleanup of the internal/native code.  Native code for
   integer and double cases have been harmonized and are now generated from a
   common code template.  This was inspired by code contributions from Peter
   Langfelder (UCLA).
 
 
Version: 0.6.3 [2013-01-13]

NEW FEATURES:

 o Added anyMissing() for data type 'raw', which always returns FALSE.

SOFTWARE QUALITY:

 o ROBUSTNESS: Added system test for anyMissing().

 o ROBUSTNESS: Now S3 methods are declared in the namespace.
 
 
Version: 0.6.2 [2012-11-15]

SOFTWARE QUALITY:

 o CRAN POLICY: Made example(weightedMedian) faster.
 
 
Version: 0.6.1 [2012-10-10]

BUG FIXES:

 o In some cases binCounts() and binMeans() could try to go past the last bin
   resulting a core dump.

 o binCounts() and binMeans() would return random/garbage values for bins that
   were beyond the last data point.
 
 
Version: 0.6.0 [2012-10-04]

NEW FEATURES:

 o Added binMeans() for fast sample-mean calculation in bins.  Thanks to
   Martin Morgan at the Fred Hutchinson Cancer Research Center, Seattle,
   for contributing the core code for this.

 o Added binCounts() for fast element counting in bins.
 
 
Version: 0.5.3 [2012-09-10]

SOFTWARE QUALITY:

 o CRAN POLICY: Replaced the .Internal(psort(...)) call with a call to a new
   internal partial sorting function, which utilizes the native rPsort() part
   of the R internals.
 
 
Version: 0.5.2 [2012-07-02]

CODE REFACTORING:

 o Updated package dependencies to match CRAN.
 
 
Version: 0.5.1 [2012-06-25]

NEW FEATURES:

 o GENERALIZATION: Now (col|row)Prods() handle missing values.

CODE REFACTORING:

 o Package now only imports the 'methods' package.

BUG FIXES:

 o In certain cases, (col|row)Prods() would return NA instead of 0 for some
   elements.  Added a redundancy test for the case.  Thanks Brenton Kenkel
   at University of Rochester for reporting on this.
 
 
Version: 0.5.0 [2012-04-16]

NEW FEATURES:

 o Added weightedMad() from aroma.core v2.5.0.

 o Added weightedMedian() from aroma.light v1.25.2.

CODE REFACTORING:

 o This package no longer depends on the aroma.light package for any of its
   functions.

 o Now this package only imports R.methodsS3, meaning it no longer loads
   R.methodsS3 when it is loaded.
 
 
Version: 0.4.5 [2012-03-19]

NEW FEATURES:

 o Updated the default argument 'centers' of rowMads()/colMads() to explicitly
   be (col|row)Medians(x,...).  The  default behavior has not changed.
 
 
Version: 0.4.4 [2012-03-05]

SOFTWARE QUALITY:

 o ROBUSTNESS: Added system/redundancy tests for rowMads()/colMads().

 o CRAN: Made the system tests "lighter" by default, but full tests can still
   be run, cf. tests/*.R scripts.

BUG FIXES:

 o colMads() would return the incorrect estimates. This bug was introduced in
   matrixStats v0.4.0 (2011-11-11).

 
Version: 0.4.3 [2011-12-11]

BUG FIXES:

 o rowMedians(..., na.rm = TRUE) did not handle NaN (only NA).  The reason for
   this was the the native code used ISNA() to test for NA and NaN, but it
   should have been ISNAN(), which is opposite to how is.na() and is.nan() at
   the R level work.  Added system tests for this case.
 
 
Version: 0.4.2 [2011-11-29]

NEW FEATURES:

 o Added rowAvgsPerColSet() and colAvgsPerRowSet().
 
 
Version: 0.4.1 [2011-11-25]

DOCUMENTATION:

 o Added help pages with an example to rowIQRs() and colIQRs().

 o Added example to rowQuantiles().

BUG FIXES:

 o rowIQRs() and colIQRs() would return the 25% and the 75% quantiles, not the
   difference between them.  Thanks Pierre Neuvial at CNRS, Evry, France for
   the report.
 
 
Version: 0.4.0 [2011-11-11]

SIGNIFICANT CHANGES:

 o Dropped the previously introduced expansion of 'center' in rowMads() and
   colMads().  It added unnecessary overhead if not needed.
   
NEW FEATURES:

 o Added rowRanks() and colRanks().  Thanks Hector Corrada Bravo (University
   of Maryland) and Harris Jaffee (John Hopkins).
 
 
Version: 0.3.0 [2011-10-13]

PERFORMANCE AND MEMORY:

 o SPEEDUP/LESS MEMORY: colMedians(x) no longer uses rowMedians(t(x)); instead
   there is now an optimized native-code implementation.  Also, colMads()
   utilizes the new colMedians() directly.  This improvement was kindly
   contributed by Harris Jaffee at Biostatistics of John Hopkins, USA.

SOFTWARE QUALITY:

 o Added additional unit tests for colMedians() and rowMedians().
 
 
Version: 0.2.2 [2010-10-06]

NEW FEATURES:

 o Now the result of (col|row)Quantiles() contains column names.
 
 
Version: 0.2.1 [2010-04-05]

NEW FEATURES:

 o Added a startup message when package is loaded.

CODE REFACTORING:

 o CLEANUP: Removed obsolete internal .First.lib() and .Last.lib().
 
 
Version: 0.2.0 [2010-03-30]

 o DOCUMENTATION: Fixed some incorrect cross references.
 
 
Version: 0.1.9 [2010-02-03]

BUG FIXES:

 o (col|row)WeightedMeans(..., na.rm = TRUE) would incorrectly treat missing
   values as zeros.  Added corresponding redundancy tests (also for the median
   case).  Thanks Pierre Neuvial for reporting this.
 
 
Version: 0.1.8 [2009-11-13]

BUG FIXES:

 o colRanges(x) would return a matrix of wrong dimension if 'x' did not have
   any missing values.  This would affect all functions relying on colRanges(),
   e.g. colMins() and colMaxs().  Added a redundancy test for this case.
   Thanks Pierre Neuvial at UC Berkeley for reporting this.

 o (col|row)Ranges() return a matrix with dimension names.
 
 
Version: 0.1.7 [2009-06-20]

BUG FIXES:

 o WORKAROUND: Cannot use "%#x" in rowTabulates() when creating the column
   names of the result matrix.  It gave an error OSX with R v2.9.0 devel
   (2009-01-13 r47593b) current the OSX server at R-forge.
 
 
Version: 0.1.6 [2009-06-17]

DOCUMENTATION:

 o Updated the help example for rowWeightedMedians() to run conditionally
   on aroma.light, which is only a suggested package - not a required one.
   This in order to prevent R CMD check to fail on CRAN, which prevents
   it for building binaries (as it currently happens on their OSX servers).
 
 
Version: 0.1.5 [2009-02-04]

BUG FIXES:

 o For some errors in rowOrderStats(), the stack would not become UNPROTECTED
   before calling error.
 
 
Version: 0.1.4 [2009-02-02]

NEW FEATURES:

 o Added methods (col|row)Weighted(Mean|Median)s() for weighted averaging.

DOCUMENTATION:

 o Added help to more functions.

SOFTWARE QUALITY:

 o Package passes R CMD check flawlessly.
 
 
Version: 0.1.3 [2008-07-30]

NEW FEATURES:

 o Added (col|row)Tabulates() for integer and raw matrices.

BUG FIXES:

 o rowCollapse(x) was broken and returned the wrong elements.
 
 
Version: 0.1.2 [2008-04-13]

NEW FEATURES:

 o Added (col|row)Collapse().

 o Added varDiff(), sdDiff() and madDiff().

 o Added indexByRow().
 
 
Version: 0.1.1 [2008-03-25]

NEW FEATURES:

 o Added (col|row)OrderStats().

 o Added (col|row)Ranges() and (col|row)(Min|Max)s().

 o Added colMedians().

 o Now anyMissing() support most data types as structures.
 
 
Version: 0.1.0 [2007-11-26]

NEW FEATURES:

 o Imported the rowNnn() methods from Biobase.

 o Created.