Version 13.1.9 2026-01-24
==========================
Fixes os-compatibility issues with previous version.

Version 13.1.8 2026-01-21
==========================
- Fixes problem in gvar.svd that arose when trying 
  to gv.dump/load results in Python 3.14.

- Minor fixes to documentation.

Version 13.1.7 2025-07-13
=========================
- Trivial bug fix in fmt_values().
- Fixes gvar.gvar(mean, cov) so it works when cov is not c-contiguous or when it is read-only.

Version 13.1.6 2025-04-03
=========================
Fixes documentation error that was causing problems.

Version 13.1.5 2025-01-31
=========================
Fixes error in pxd files introduced by last update.

Version 13.1.4 2025-01-31
=========================
Cleaned up Cython code. Updated __init__.pxd no longer cimports numpy.
It now also includes msum_gvar and wsum_gvar (for lsqfit).

Version 13.1.3 2024-12-05
=========================
Fixes incompatibility  with numpy2 in gvar.bootstrap_iter.

Version 13.1.2 2024-11-23 
==========================
Fix to pyproject.toml.


Version 13.1.1 2024-07-06
==========================
Modified so that it works with numpy>=2.0.

Version 13.1 2024-05-21
========================
- Improvements to gvar.PDFStatistics, including new keywords 
  fits, errorbars for PDFStatistics.plot_histogram.

- gvar.PDFHistogram deprecated.

Version 13.0.1 and 13.0.2 2024-03-05
=====================================
- Fixes default arguments for sample and raniter.

- Move gvar.pxd to correct place.


Version 13.0 2024/03/04
========================

- gvar.PDF is completely rewritten to increase functionality for use by the 
  vegas and lsqfit modules. The changes are incompatible with previous versions.

- Substantial extensions to gvar.PDFStatistics for use in lsqfit.vegas_fit. Adds 
  a fit to a continuous split-normal distribution to the histogram analysis. 
  Introduces an incompatibility because the attributes of a PDFStatistics 
  have different names. Adds plotting capability.

- bd = gvar.BufferDict(d, keys=[k1, k2, ...]) copies d[k] into bd[k]
  for every k in keys that is also in d. The results are now copied in the 
  order specified by keys. 

- Fixes a bug in gvar.BufferDict.all_keys that in certain situations included 
  the same key twice.

- Adds BufferDict.lbatch_buf and BufferDict.rbatch_buf for extracting batch buffers 
  from batch BufferDicts.

- Fixes bug in gvar.BufferDict.all_keys which caused crashes when keys 
  were not strings.

- Fixes inconsistencies in gvar.SVD when svdcut=0.0. 

Version 12.1 2023-12-04
========================
- Adds batch capability to gvar.sample and gvar.raniter so samples can 
  be created in large batches, which is much more efficient if large 
  numbers of samples are needed. Also adds similar capability to 
  gvar.BufferDict when making clones. These changes are useful
  for (fast) Monte Carlo simulations: see the section on Sampling GVars 
  in the tutorial.

- Updates use of numpy's random module to conform with changes in 
  numpy version 1.17. See the documentation for :func:`gvar.ranseed`.

- New case study (on Creating an Integrator) that illustrates how to 
  repurpose an existing library function to accommodate GVars.

- The import time for gvar is greatly reduced, especially if numpy has 
  already been loaded. 

- Simplifies gvar.__version__ (no longer uses importlib.metadata, which is buggy
  and increases gvar's import time significantly).
  
- Minor fixes to GVar.set, GVar.__format__, evalcov, evalcov_blocks, 
  evalcov_blocks_dense, and gvar.dataset.avg_data (when median=unbias=True).

Version 12.0 2023-11-01
===========================
Introduces more modern string formatting for GVars, including support for 
the e, f and g formats used to format floats. For more details see the section 
on "Formatting GVars for Printing and Display" in the GVar documentation.
The default format is somewhat different from before.

Version 11.11.17 2023-10-12
============================
Fixes obscure bug on Windows machines when evalcov_blocks(g) is called with a 
single GVar g. The tests should now work on Windows machines.
 
Version 11.11.16 2023-10-11
============================
Identical to previous version but now offers Python wheels for installation.

Version 11.11.15 2023-09-27
============================
Tweak to the new build instructions introduced in 11.11.14.

Version 11.11.14 2023-09-27
=============================
- Adds numpy.import_array() to Cython files for Cython3. 

- Changes the build instructions in pyproject.toml to deal with 
  incompatibilities between different versions of numpy.

- Adds unbias option to dataset.avg_data().

Version 11.11.13 2023-08-27
=============================
Bug fix to forbid distribution function names in BufferDict.add_distribution() 
that contain parentheses (which would cause problems if allowed).

Version 11.11.12 2023-07-22
===========================
Bug fixes to make code compatible with Cython v3.0 and numpy v1.25.

Version 11.11.11 2023-06-11
============================
- Minor error in gvar.powerseries.PowerSeries.__truediv__ that triggered 
  exception when it shouldn't have.

- Fixes bug that caused a crash when old pickle files containing GVars 
  were read in. Note that pickling GVars directly is generally not 
  a great idea; use gvar.dump and gvar.load instead of pickle.load/dump.

Version 11.11.10 2023-05-11
===========================
- New and much simpler installation code. Works with Python3.11, even if 
  numpy is not installed.

- Fix readthedocs setup.

Version 11.11.2 2023-02-26
==========================
Cleans up inconsistent treatment of data types by gvar.BufferDict. 

Version 11.11.1 2022-11-11
===========================
- Fixes 2 minor bugs in gvar.dump/load. 

Version 11.11 2022-11-02
==========================
Mostly bug fixes.

- Fixes bug causing gvar.var to crash under certain circumstances 
  when one or more arguments are not a GVars.

- Fixes pickling of GVars using pickle.dump/load. Added warning that 
  using pickle.dump/load loses the correlations between GVars.
  To preserve correlations use gvar.load/dump instead.

- Fixed bug in pickling of BufferDicts (didn't capture distributions properly).

- Forces Cython to regenerate *.c files when using Python 3.11 or later (deals
  with incompatibilities introduced by 3.11).

- gvar.load/dump add support for classes with __slots__.


Version 11.10.1 2022-08-01
===========================
- Fixes bug that prevented gvar.gvar(x, cov) from using 
  a read-only covariance matrix cov.

- Fixes a bug in BufferDict.uniform which allowed 
  incorrect arguments to cripple BufferDict. The 
  incorrect arguments now raise exceptions.

Version 11.10 2022-05-12
=========================
Adds the possibility of phasing the construction of a set 
of correlated GVars. gvar(ymean, ycov, x, xycov) returns 
primary GVars y that have covariance matrix ycov and 
covariance matrix xycov with previously defined 
primary GVars x. The first and third arguments 
(ymean and x) are 1-d arrays; the other two are 
2-d arrays.

Version 11.9.7  2022-05-01
==========================
Thanks to Giacomo Petrillo for these ideas and issues.

- Fixes bug in raniter when svdcut<0.

- Fixes var, sdev, evalcorr when given [] as an argument -- ie,
  they no longer crash.

- More legible __str__ and __repr__ for BufferDict. Thanks to 
  Giacomo Petrillo for the code.

- New iterator all_keys() for BufferDict which iterates over 
  the dictionary keys and the extension keys. Thanks again to 
  Giacomo Petrillo for code.

Version 11.9.6 2022-03-13
==========================
Fixes bug in BufferDict.del_distribution that caused it to crash. Also 
adds more protection to BufferDict.add_distribution. Adds new 
static method BufferDict.has_distribution.

Version 11.9.5 2021-12-13
===========================
- Fixes issue with Python 3.10.

- New keyword (keys) in gvar.tabulate. 

Version 11.9.4 2021-09-06
==========================
Bug fix in evalcov_blocks for problem in Windows installations.

Version 11.9.3 2021-08-14
==========================
- Very minor bug fix to gvar.erf and new implementation.

- Bug fix for graphics in PDFHistogram.

- Minor optimizations.

Version 11.9.2 2021-04-15
=========================
- Bug fix in evalcov to get rid of (unusual) divide-by-0 error.
  Thanks to Giacomo Petrillo for pointing out the problem.

- New keywords for correlate: upper, lower, verify. These 
  make it easier to build the correlation matrix from an 
  upper or lower matrix, and to verify (optionally) the 
  final GVars.

Version 11.9.1 2020-11-25
=========================
Yet another tweak to gvar.dataset.avg_data(dataset). This makes 
it a bit faster for very large problems and avoids a bug 
in scipy's sparse matrix code when the covariance 
matrix has more than 2**31 elements.

Version 11.9 2020-11-22
========================
- Further fixes to gvar.dataset.avg_data(dataset). The previous 
  fix (in v11.8) could lead to problems for dictionary datasets 
  where entries dataset[k] have different sample sizes. The new 
  version has a new keyword (mismatch) that offers three options 
  for dealing with this situation.

- Several functions (eg, gvar.var) generated segmentation errors if GVars
  in the arguments were replaced by None. This has been fixed. What happens
  in such cases depends on the function, but segmentation faults should 
  not happen.


Version 11.8 2020-11-15
=======================
Small upgrade to support feature request: dataset.avg_data(dataset) no longer 
discards data when dataset is a dictionary whose entries have different 
sample sizes. It used to truncate entries that had more samples (by 
discarding from the end) so that all entries had the same (smallest)
sample size. 

Also added keyword uniform to raniter.


Version 11.7 2020-07-17
========================
- Adds support for multivariate power series to gvar.powerseries, as well 
  as minor bug fixes. Adds error function (erf(powerseries)).

- Adds new numerical analysis module for Pade approximants of functions.

Version 11.6 2020-06-11
========================
- gvar.regulate is a new tool for regulating singularities 
  in correlation matrices. It has two regulation schemes, 
  one tailored for Cholesky decompositions of the 
  correlation matrix (eps), and the other using an 
  SVD decomposition (svdcut). The Cholesky/eps scheme is 
  typically a bit faster, but is less robust numerically. 
  The SVD scheme uses gvar.svd. The Cholesky implementation 
  is based on a recommendation from Giacomo Petrillo.

- Keyword add_svdnoise in gvar.svd has been replaced by keyword noise. 
  The old keyword still works, however (for legacy code).
  
- gvar.chi2, gvar.raniter, gvar.bootstrap_iter now have the option
  of specifying either eps or svdcut (passed on to gvar.regulate).

- More efficient implementations of gvar.sdev and gvar.var. 

Version 11.5.2 2020-05-09
==========================
Fix documentation problems introduced in 11.5.1.

Version 11.5.1 2020-05-09
==========================
More tuning of evalcov, for better performance on dense matrices of 
size 15x15 or larger. Also improved doc-strings for Cython-generated 
functions.

Version 11.5 2020-05-06
========================
Significant performance upgrades to evalcov and evalcov_blocks, 
especially for large systems.

- Fixes an inefficiency in gvar.evalcov_blocks that was introduced in v11.4.

- New optimization in gvar.evalcov for large, dense covariance matrices. 
  The improvement can be as large as a factor of 3-4 speedup for matrices
  larger than (roughly) 200x200. Suggested by Giacomo Petrillo. 


Version 11.4 2020-04-29
=======================
Significant rewrite of gvar.evalcov_blocks(g), making it much 
faster (thanks to Giacomo Petrillo for the suggestion). Also 
gvar.gvar(x,cov) has new keywords: verify and fast. See 
documentation. 

Version 11.2 2020-03-18
========================
Minor fixes and updates.

- Fixes (substantial) inefficiency in gvar.ode.Integrator. This was a coding 
  error that gave correct results but sometimes at a cost much larger than 
  necessary. Adds options hmax, maxstep to ode.Integrator.

- Adds gvar.abs. This is more less same as gvar.fabs, but unlike fabs works 
  for something like np.array([1., -2.], dtype=object). Treatment of GVars is
  the same. 

Versions 11.1 2020-03-01
==========================
Further upgrade to gvar.dump and gvar.load to allow 
user-defined classes some control over how GVars 
are removed and distributed. New functions are 
available to help: gvar.remove_gvars and 
gvar.distribute_gvars. These changes respond
to a feature request for the lsqfit module.

Version 11.0 2020-02-25
========================
Significant upgrade to gvar.dump and gvar.load, which now can 
handle data structures that contain mixtures of GVars with 
other data types, and that can include nested dictionaries, lists,
tuples, etc. The old versions (v10.1.2) of these 
are still available as gvar.gdump and gvar.gload. 

The new versions of dump/load are much more flexible, but 
unlike the old versions only work with pickle (not json) 
as the serializer; there is no longer a "method" keyword.
gdump/gload still offer a choice between pickle and json.

Other additions: gvar.dependencies, gvar.missing_dependencies,
and gvar.filter.

Version 10.1.2 2020-02-24
==========================
Fixes bug introduced into dump in version 10.1 (with thanks to 
Andre Walker-Loud for pointing it out).

Version 10.1.1 2020-02-23
=========================
Fixes minor bug/inefficiency in gvar.dependencies and gvar.dump.

Version 10.1 2020-02-22
=========================
Significant improvement to gvar.dump/load and some new functions. 
All changes are backwards compatible. Note that this code 
still works with Python 2.7 (as well as Python 3.X, of course)
but it seems likely this will not continue far into 
the future, as developers for the several packages
used by gvar abandon Python2, now that it is formally
discontinued.

- New versions of gvar.dump and gvar.dumps that can preserve 
  dependencies between derived and primary GVars (to allow, for 
  example, the construction of error budgets). Update for 
  pickling of gvar.BufferDicts.

- New function gvar.qqplot() for making QQ-plots.

- Slight modification of gvar.chi2() to support 
  implementation of gvar.qqplot().

- New function gvar.is_primary(g) to determine whether GVars
  are primary or derived. Also gvar.GVar.is_primary(). Also
  new function gvar.dependencies for collecting relevant 
  primary GVars.

- Function gvar.deriv(g, x) now works for an array x of GVars 
  (not just a single GVar). Also gvar.GVar.deriv(x).

- Fixed bug (that caused crashes) in (undocumented) low-level 
  gvar.gvar() usage. 

Version 10.0.3 2020-02-07
=========================
Changed default string representation used when 
printing GVars so that it always shows all 
significant digits in the mean. This was 
not the case before when answers were 
very accurate. 

Version 10.0.2 2020-02-03
=========================
Bug fix so that gvar.cspline.CSpline works with only 2 knots. 

Version 10.0.1 2020-02-03
=========================
Bug fix for Windows users (thanks to Christoph Gohlke).

Version 10.0 2020-02-02
=======================
New features and improvements to existing features. 
There are incompatible changes in gvar.loads and 
gvar.cspline.CSpline, though most old code should
still run.

- gvar.cspline.CSpline has been reimplemented. It now supports 
  different algorithms (select using keyword alg) and the 
  default algorithm is a monotonic spline rather than a 
  classical cubic spline.

- Adds support in gvar.BufferDict for uniform distributions: see 
  gvar.BufferDict.uniform. This is for use by lsqfit.

- gvar.load() and gvar.dump() use a new protocol (for serializing
  GVars) that make them substantially more efficient for large
  problems. gvar.load() can read files created with the 
  old protocol, but if you want to update the format used 
  in an old file, load it and dump it again:
  dump(load('filename'), 'filename'). gv.loads() no longer
  has a method keyword since the method is inferred from 
  the data.

- Pickle format used for BufferDicts is changed but the new code
  can handle data from pervious versions.

- New option (compress) for gvar.evalcov_blocks().


Version 9.2.1 2019-12-04
========================

- Fixes bugs in gvar.load and gvar.dump caused by recent security upgrades
  to pyYAML.


Version 9.2 2019-05-08
========================

- gvar.dataset.svd_diagnosis has new keyword process_dataset that
  allows it to be used in more general contexts.

- Fixes bug that prevented Dataset from reading compressed files
  in Python 3.


Version 9.1 2019-03-19
=========================
Minor update.

- Simple arithmetic is now supported for BufferDicts having the same layout.

- Much improved the heuristic used in gvar.dataset.svd_diagnosis to suggest
  an svdcut. The new heuristic gives lower estimates usually. Also has a new
  option for the input data. And fixed bug that occasionally caused crashes.

- Adds new option (add_svdnoise) to gvar.svd. This is primarily for use
  by lsqfit. Also modifies the return value to contain information related
  to the SVD cut (dof, nblocks, ...) (instead of storing that information in
  gvar.svd itself, which is a bad idea).

- New convenience function gvar.sample(g), which is the same as
  next(gvar.raniter(g)).


Version 9.0.3 2018-10-03
========================
Very minor update.

- Fixes small inefficiency in installation script.

- Small optimization for cloning BufferDicts.

- load and loads now able to load files dumped using versions of
  gvar older than 8.4. The old format is different.

Version 9.0.2 2018-08-03
========================
- Fixes bug in evalcov_blocks that missed out correlations in certain
  unusual situations. This affects svd.

- Fixes incompatibility with latest numpy.

Version 9.0.1 2018-08-01
=========================
Fixes bug in BufferDict that made certain types of dictionary key fail.

Version 9.0 2018-07-27
=========================
Introduces new BufferDicts that are mostly but not completely compatible with
the old. The incompatible differences are in obscure features that were mostly
unused.

- BufferDicts reimplemented to give them more conventional/consistent
  dictionary behavior. Also have added functionality to support
  (limited) non-Gaussian distributions. Better documentation for
  BufferDicts.

- Internal data type ExtendedDict is now gone because BufferDict can
  provide the needed functionality in a much more robust (and simple)
  fashion.


Version 8.5.1 2018-07-21
========================
Very minor fixes.

- Small improvement to gvar.chi2(...).

- Default svdcut changed to 1e-12 everywhere.

Version 8.5 2018-07-21
======================
Very minor enhancement.

- gvar.linalg.lstsq now supports both weighted and unweighted least squares
  for solving a @ x = b when b has a covariance matrix.

Version 8.4 2018-07-15
======================

- New functions eigh, svd, lstsq added to gvar.linalg. Also have a faster
  version of solve.

- gvar.load and gvar.dump now work with yaml as well as with pickle and json.

Version 8.3.7 2018-07-08
========================
Bug fixes.

- Bug fix in evalcov_blocks() that affects some Windows installations
  (mismatched integer types in cython code).

- Bug fix for ode to deal with 0(0)/0(0) fringe case.

- Minor documentation fixes.

- Bug fix in GVar.__richcmp__ for Python 2.7 on macOS 10.13.5.

- Default svdcut=1e-15 added to raniter. Having a small non-zero
  svdcut is standard elsewhere in code.

- Improved handling of cython in makefile.


Version 8.3.6 2018-03-18
============================
Minor update to keep up with changes in numpy.


Version 8.3.5 2018-02-15
========================
- More improvements for setup.py.


Version 8.3.4 2018-02-15
=========================
- New setup.py that: a) does not require cython; b) works even if
  numpy is not installed (by installing numpy before it is required).
  Previous versions of setup.py did not work unless cython and numpy
  were already installed. This version uses setuptools to install
  numpy and, optionally, cython before they are used to build gvar.
  cython is not used by default; to turn it on set USE_CYTHON=True in
  setup.py.

- Fixed formatting bug in gvar.fmt_errorbudget().

- Modifications to support hosting documentation by ReadTheDocs.

- Minor bug (error message) in gvar.Histogram.


Version 8.3.3 2017-08-03
=========================
Fixes typo in gvar.correlate that makes gvar compilation fail on some
machines.

Version 8.3.2 2017-06-22
=========================
json support in gvar.dump, gvar.load, etc was not quite right in v8.3.1 and so
was unreliable. The situation is improved here. Note that these functions are
meant to extend and improve on gvar.BufferDict.dump, etc which functions are
still available but will get limited support going forward.

Version 8.3.1 2017-06-16
==========================
A very minor updata:

- gvar.dump, gvar.load, etc can now use json instead of pickle
  for serializing collections of GVars. pickle is probably more
  robust.

- bug fix in evalcorr to deal with very rare condition that caused
  the routine to crash (condition caused by std dev = 0.0 for a GVar).

Version 8.3 2017-06-03
======================
Bug fixes and a little bit of new functionality.

- A new utility gvar.dataset.svd_diagnosis() analyzes a dataset
  containing Monte Carlo results for multiple variables to determine
  whether or not an SVD cut is needed for the correlation matrix
  describing these variables. The smallest eigenvalues of a correlation
  matrix can be badly underestimated when the number of
  Monte Carlo samples is insufficiently large. This makes the correlation
  matrix more singular than it should be, which can lead to problems,
  for example, when trying to fit the correlated data. This utility
  estimates the size of the svd cut to use on the correlation matrix.
  A new case study illustrates its use.

- Fix plotting by PDFHistogram.make_plot to account for (incompatible)
  change introduced in matplotlib-2.0.0's pyplot.bar.

- gvar.dataset.bin_data(xx) returns a gvar.dataset.Dataset ordered
  dictionary if xx is a dictionary.

- pickle now works for (the mostly hidden classes) gvar.svec and gvar.smat
  and therefore for x.internaldata where x is a GVar. Tricky to use
  internaldata, however.


Version 8.2.3  2017-04-30
==========================

- Slightly faster memory allocation for GVars.

- Fixed problems in gvar.BufferDict caused by changes in Python's
  dictionaries starting at Python 3.5. The new code has been tested
  (and should work) with Python 3.4, 3.5 and 3.6.

Version 8.2.2 2017-03-17
=========================
Minor/rare bug fixes:

- SVD(mat, rescale=True) works now when mat has a zero on the diagonal. Such
  a zero is a very unusual occurrence so this fix will have no impact
  on almost all codes. This repairs a bug introduced in v8.1.

- GVars whose mean or standard deviation are nans will no longer cause
  an exception when they are printed. Thanks to Daniel Hackett.

Version 8.2.1 2017-02-11
=========================

- Fixed bug in gvar.gvar that mishandled GVars with standard
  deviation exactly equal to 0, when they were in arrays. This
  caused segmentation faults occasionally.

- Tiny fix to Dataset for use with hdf5 files.

Version 8.2 2017-02-06
========================
Very small changes, mostly to support other modules.

- gvar.dataset.Dataset can now read from files that use hdf5 format.

- del g[k] and g.get(k) now implemented for any BufferDict g.

- Can specify buffer data type when creating a BufferDict from
  a dictionary or list of tuples: g = BufferDict(d, dtype=object).

- Fixed small bug in gvar.tabulate (headers=False didn't work).

Version 8.1 2017-01-09
===========================
Mostly internal changes needed by other code.

- Improved internal treatment of sparse covariance matrices in
  gvar.svd (more sparse now). This involved improvements in svec and smat,
  which now look for and remove entries that are 0; and it involved a new
  function evalcov_blocks that constructs the non-zero block diagonals of the
  covariance matrix (as opposed to the entire matrix, which is what evalcov
  does). This should lead to small speedups but also greatly reduce memory
  consumption in cases where evalcov was used to construct covariance matrices
  for large numbers of GVars (eg, lsqfit).

- Two new convenience functions: cov(g1,g2) which is evalcov([g1,g2])[0,1]
  and corr(g1,g2) which returns evalcorr([g1,g2])[0,1].

- gvar.ranseed now accepts single integers, as well as lists of
  integers, as an argument.

- New low-level routines gvar.disassemble and gvar.reassemble
  that should be used with care.

- changed gvar.chi2() so its output can be formatted with
  gvar.fmt_chi2().

Version 8.0 2016-11-05
=======================
This is a small change but one that is not compatible with previous
versions. I accidently introduced a circular import when adding
PDFIntegrator to gvar. PDFIntegrator has to be part of the vegas
module, not the gvar module. So gvar.PDFIntegrator needs to be
replaced by vegas.PDFIntegrator in existing code. This is the only
change and is irrelevant if you are not using module vegas together
with gvar (or lsqfit).

There are also some internal changes to correct data-type mismatches
on windows machines.

Version 7.3.1 2016-08-17
==========================
This is a very small change to the new PDFIntegrator class (new as of
v7.3) that results in much improved performance for very high dimensions
(eg, integrating over 100 variables). PDFIntegrator has a way of
pre-adapting the integrator during initialization, so that it is almost
fully adapted to the PDF. This has little impact for low-dimension
integrals (eg, 3 variables) but can make a huge difference for
very-high-dimension integrals, causing them to work for much smaller
values of neval. This feature is actually in the v7.3 code but was
accidentally left unactivated. This is merely an optimization.

The only other correction is a small, behind-the-scenes improvement
in PDFStatistics.

Version 7.3 2016-08-14
========================
New features include PDFIntegrator for integrating over PDFs and the
ability to read compressed files in gvar.Dataset. Several bug fixes as well.

- New class PDFIntegrator for evaluating expectation values weighted
  by the probability density function (PDF) for arbitrary multi-dimensional
  Gaussian distributions. This class uses the vegas module to evaluate
  the multi-dimensional integrals, and optimizes the integrands for vegas.
  vegas uses adaptive Monte Carlo integration and so can handle fairly
  high-dimension integrals (dim=10, 20, 50 ...) efficiently. The vegas
  module must be installed separately. Three other modules are provided
  in addition: PDF, PDFStatistics and PDFHistogram. See the section
  on "Non-Gaussian Expectation Values" in the tutorial for an example.

- gvar.Dataset can now read gzipped and bzipped files. It looks for .gz
  and .bz2 at the ends of file names to identify which files are
  compressed.

- Made gvar.gvar(a, sd) substantially faster (30%) when a is a large array.
  Replaced a Python list by a numpy array -- corrects an oversight from
  before. This probably won't have a big effect on most codes since
  gvar creation is usually a small part of the cost.

- Bug fixes in gvar.powerseries that make it work better with
  coefficients that are numpy arrays.

- Bug fixes in gvar.cspline when spline is used outside of its range,
  with extrap_order set to 0, 1 or 2. The array size returned
  by the spline function (or derivatives) was not always correct.
  Additional test code.

- Small change in BufferDict to help with legacy code issues.

- BufferDict(d) for d=dictionary no longer stores keys in sorted order. This
  never made sense but also causes trouble in python3 when keys of mixed type
  are used.

Version 7.2  2015-08-10
=======================

- Fixed error in evalcov(g) and evalcorr(g) when g is a dictionary.
  The return values are doubly indexed dictionaries c[i,j] whose
  values correspond to g[i] and g[j]. The shape of c[i,j] is
  an array whose shape is the sum of the shapes of g[i] and g[j],
  where scalars are treated as shape=(1,) objects. Also evalcorr(g) worked
  incorrectly when elements g[k] were arrays.

- Added correlate(g, corr) which takes a set of uncorrelated GVars in g
  and adds correlations as specified by correlation matrix corr. It is
  common for people to publish correlated data as a set of means and
  standard deviations, plus a correlation matrix. This routine
  facilitates converting such data into GVars.

- Minor corrections to documentation and improved testing for evalcov
  and evalcorr.


Version 7.1 2015-07-24
=========================

- Added gvar.tabulate(g) for tabulating values stored in arrays or
  dictionaries of GVars.

- Fixed typos in the tutorial.

- Documentation fix in gvar.BufferDict.

Version 7.0.3 2015-06
=========================

- fixed default initialization of gvar.ranseed so it doesn't generate
  errors due to new restrictions on numpy's random generator.

Version 7.0.2 - 2015-06-17
===========================
Changed to pip + distutils for installation. Fixed inconsistencies in
INSTALLATION.txt.

Version 7.0.1 - 2015-06-16
==========================
Had to replace setuptools with distutils again, because I could not get cython
pxd files to work properly with the eggs, causing build problems for
lsqfit.

Version 7.0 - 2015-06-16
==========================

The biggest change in this version is that gvar is now distributed
independently of lsqfit; it will no longer be bundled with the latter.  This
is better because gvar is used by packages other than lsqfit.  There are also
two very minor but incompatible changes, in the interfaces for gvar.chi2 and
gvar.ode, that require the major version number to increase from 6 to 7
(because of  semantic versioning). Finally there are some significant
additions to the numerical analysis routines provided with gvar.

- The first incompatible change concerns a rarely used feature
  of gvar.chi2. Previous versions stored information about dof and Q in
  as attributes of the function; these are now stored in the result of the
  function, which is much more natural.

- The second incompatible change is in gvar.ode. The behavior when the
  step size is reduced below hmin is changed. Before an exception was
  issued and execution stopped. Now a warning is issued, the step size
  is not reduced and execution continues. The warning is necessary since
  it means that errors may not be under control to the extent hoped.

- Added gvar.ode.integral for 1d integrals.

- Added two new numerical analysis packages to gvar:

    gvar.root - finds roots of 1d functions
    gvar.linalg - basic linear algebra (det, inv, solve, eigvalsh ...)

  As for the other numerical analysis methods, the new methods work with any
  combination of numbers and GVars.

- gvar.cspline now allows you to specify what order polynomial is used for
  out-of-range points. In the past a cubic polynomial was used, based on
  the value and derivatives at the knot nearest to the out-of-range point.
  Now one can specify a lower-order polynomial by setting parameter
  extrap_order equal to 0, 1, or 2.

- Added gvar.equivalent(g1, g2, rtol, atol) which checkes whether or not
  the GVars in g1 and g2 are the equivalent (ie, g.mean and g.der agree
  for each GVar). This is useful for debugging purposes.

- Installation will now use setuptools instead of disutils, provided
  setuptools is installed. setuptools automatically ensures that
  modules needed by gvar are installed and sufficiently
  up-to-date. If setuptools is not installed, distutils is used and
  dependencies are not checked --- this must be done by hand.

- Tiny addition to gvar tests.




Version 6.0 - 2014-11-01
========================

This upgrade deals with some issues created by the most recent update  to
numpy (v1.9.0). The verion number increases to 6 because a (very minor) part
of the gvar.powerseries interface had to change (and therefore by the rules of
semantic versioning one must  increase the major verion number). BufferDicts
are also subtly changed, though in a way that shouldn't affect much code.
A couple of enhancements are included as well.

- gvar.BufferDicts are now derived from collections.OrderedDict. This is
  a consequence of the numpy change, but it is probably a good idea anyway.
  The interface is the same but pickling has changed. This means that
  BufferDicts pickled with the old software cannot be unpickled with
  the new. A very minimal module, oldbufferdict.py, is included in lsqfit
  for converting old pickled data to the new format; see documentation
  inside the file. Data stored using json is unaffected.

- The power series coefficents of gvar.powerseries.PowerSeries variable
  p are now accessed via attribute c: p.c[0], p.c[1] etc. Formerly one
  could access the coefficients using array notation --- p[0] for p.c[0],
  etc --- but this no longer works with numpy. The old notation was a bad
  idea in any case.

- gvar.fmt_errorbudget has a new option called verify. Set verify=True
  to have the code check that the error budget is complete and has no
  double counting of errors; a warning is issued if this is not the case.

- Added arctan2 to the list of functions that work with GVars,
  using code from Matt Wingate. (NB, the documentation shows how to
  create additional functions (eg, erf(x)) using gvar_function.)

- Obscure bug fix in gvar.ode.Integrator allows GVars as the
  independent variables.

- Rearranged documentation now includes case studies. These are self
  contained examples that are extensively annotated (and relatively simple).
  There are only two at the moment but more are forthcoming.


Version 5.0.1 - 2014-09-13
===========================
This is a very minor bug fix.

- Modified gvar.mean(g), gvar.sdev(g), gvar.var(g) so that g can
  include elements that are not GVars. These elements are assumed to
  have zero error, so their value is their mean. This means, for
  example, that g can be all GVars or all floats or a mixture.

- Fixed bug in gvar.ranseed which caused it to crash under python3 in
  certain circumstances.

Version 5.0 - 2014-06-26
=========================
A minor upgrade to provide functionality for other libraries. The first
change is not backwards compatible, but is unlikely to break much or any
code (except mine).

- fit.svdcorrection is now the sum of all svdcorrections, not a vector. The
  sum is all that is needed to construct error budgets, which is the
  only use for fit.svdcorrection. The sum is easier to manipulate -- for
  example, two svdcorrections can now be added to give a composite
  svdcorrection (as opposed to concatenated before).

- Fixed innocuous bug in lsqfit.nonlinear_fit.format().

- gvar.dataset.Dataset.slice can now apply slices to multiple dimensions.

- Added __format__ to gvar.GVar. This converts the GVar to a string
  and then applies the format to the string. This was needed to make
  python3 and python2 function similarly in certain situations.

- More polishing of the documentation (especially about svd).

- Beginning with this version of lsqfit, version numbers will be consistent
  with the rules of "semantic versioning" (http://semver.org). The most
  important consequence is that any change that is *not* backwards compatible
  with the current interface (as defined by the documentation) will be signaled
  by a change in the major version number, where the full version number has the
  format major.minor.patch (so the major version number is 5 for version 5.0).
  A side effect of this is that increases in the major version number can
  be triggered by relatively obscure changes in the interface that will have
  no effect whatsoever on almost all existing codes. The current move to 5.0
  is an example.

Version 4.8.5.1 - 2014-05-30
============================
This release is almost identical to the previous release. It is mainly to do
with getting the software registered properly online.

- Small improvement to the user interface in gvar.ode. It now works with
  vectors of x values, returning vectors of y values.

Version 4.8.5 - 2014-05-29
==========================

- Bug fix so recent changes in lsqfit don't cause crashes on Windows systems.

Version 4.8.4 - 2014-05-23
===========================
Small bug fixes.

- gvar.dataset.Dataset objects now work with Python's pickle. Also
  Dataset objects can now be initialized from other Dataset object, or from
  dictionaries, or lists of key-data tuples (like dictionaries).

- lsqfit.wavg now works with priors.

- Removed calls to C's lgamma function because it isn't available to
  old C compilers. Use Python's math.lgamma instead, which should always be
  available.

Version 4.8.3 - 2014-05-21
==========================
Bug fix: the new svd system didn't work with older versions of numpy (earlier
than 1.8). It now does. The problem was in lsqfit; gvar was fine.
There was a related issue in gvar.cspline which is also fixed.

Version 4.8.2 - 2014-05-16
==========================
Very small additions to facilitate pickling and other serialization of GVars. Changed
gvar.evalcov and gvar.gvar so that newg = gvar.gvar(gvar.mean(g), gvar.evalcov(g))
for some array or dictionary g containing GVars creates a new array/dictionary
with the same means and covariance matrix. Added better documentation
on the gvar page about pickling gvars, and also added new functions
(dump, dumps, load, loads) for pickling collections of GVars.

Version 4.8.1 - 2014-05-14
===========================
Bug fix.

- fixed negative svdcuts (which one shouldn't use anyway), which were broken.

- gvar.cspline.CSpline functions now return scalars rather than 0-D numpy
  arrays when given scalar arguments.

Version 4.8 - 2014-05-12
========================
This release has a reworked framework for SVD cuts. The new framework is
simpler, but also more robust --- SVD cuts are applied separately
to each block-diagonal sub-matrix of the data-prior covariance matrix. This
is more accurate, faster, and uses less memory, though the differences
aren't substantial for most fits.

- svdnum is no longer an option for lsqfit.nonlinear_fit, raniter, etc.
  This parameter makes little sense in the new framework. Make svdcut
  negative if it is important to get rid of certain eigenmodes. (Generally
  it is better to keep the modes, but with larger variances. This is
  what happens when svdcut is positive.)

- New information included in the Settings part of a fit report:
  number of eigenmodes modified by SVD cut, time needed for fit.
  The number of iterations is moved from the top line of the report
  to the last line, with the time.

- Better notation for lsqfit.transform_p (mostly compatible with old
  usage).

- Added a new numerical analysis routine (joining gvar.ode and gvar.cspline)
  in gvar for manipulating power series representations of functions:
  gvar.powerseries. The numerical analysis routines in gvar have been
  adapted so that they work with GVars and/or floats. More will be added
  in the future.

- Fixed bug in gvar.cspline (derivatives not computed correctly); added
  better tests for cspline.

- Fixed documentation errors in gvar.dataset.

- gvar.dataset.avg_data(s) for a dictionary s now (optionally) issues
  a warning if different entries in s have different sample sizes. The
  smallest sample size is chosen for all entries in such situations.

- Added option to gvar.ode.Integrator to allow custom adaptive
  strategies. Also offers a different interface to the integrator.

- Fixed minor irregularity (not really a bug) in GVar.fmt:
  x = gvar(0.999,0.999) implied that x.fmt(2) gave 1.00(100) which
  is correct but non-standard (should be 1.00(1.00) and now is).

Version 4.7
==============
A minor upgrade but with one backwards incompatible change (that is unlikey
to affect much code). The incompatibility is the main reason for calling this
release version 4.7 rather than 4.6.2.

- GVars can be used with any comparison operator now (==, !=, >, >=, <, <=).
  The comparison is made with the mean value of the GVar, ignoring the
  standard deviation. This is to facilitate the use/design of numeric functions
  that are designed to work with either floats or GVars. The definition of ==
  and != has changed from the past. Before x==y for two GVars was true only if
  both their means and standard deviations agreed; now it is true if their
  means agree, but their standard deviations disagree.

- Can use keyword keys to specify the which keys to use and their order when
  creating a BufferDict from a dictionary-like object.

- Significantly more documentation for the gvar module, and some rearrangement
  and expansion of the overview documentation.

- New function gvar.gvar_function for creating new arithmetic Python
  functions that can accept GVar arguments.

- gvar.gvar(x, xcov) now works when x and xcov are dictionaries. This allows
  you to deconstruct a dictionary of GVars using gvar.mean and gvar.evalcov,
  and then to reconstruct the dictionary using gvar.gvar: e.g.,
  gvar.gvar(gvar.mean(g), gvar.evalcov(g)) rebuilds g when g is an array
  of GVars, or a dictionary of GVars and arrays of GVars.

- gvar has two new numerical analysis modules: gvar.ode for integrating
  ordinary differential equations, and gvar.cspline for creating cubic
  splines from 1-d data. These modules are adapted from standard routines
  so they work with GVars as well as floats, which is why they are included
  here. Additional modules will be added as the need arises.

- When printing an lsqfit fit using fit.format(), setting option maxline=True
  (rather than a number) now causes all data to be printed.

- Other minor rearrangements and clarifications in setup.py and a few other files.

Version 4.6.1 - 2014-02-02
===========================
Cleaning up some small bugs introduced with the new lsqfit.wavg. Also
introduced an approximate but potentially much faster *fast* mode for it.

Version 4.6 - 2014-01-30
========================
The main change here is an upgrade to lsqfit.wavg.

- Somewhat incompatible change in lsqfit.wavg: When averaging arrays or dicts,
  wavg used to ignore correlations between different elements of the
  array or dict. The new wavg takes account of all correlations between
  different pieces of input data. wavg returns a GVar if averaging
  a list of GVars, a numpy array of GVars if averaging a list of arrays
  of GVars, and a Bufferdict of GVars or arrays of GVars if averaging
  a list of dicts. In each case the return value has extra attributes:
  chi2, dof, Q, time, fit. The function itself also has these attributes,
  coming from the last fit.

- gvar.mean(g) now returns g unchanged if g contains objects of type
  other than GVar. This is useful for writing functions that must work
  with either GVars or floats as arguments: gvar.mean can be used to
  strip the sdev off of GVars where it isn't needed or wanted.

- New function gvar.asbufferdict(g) converts dictionary g to a
  BufferDict unless it already is one, in which case it returns g.
  The keys in the final result can be restricted by adding a
  a list of keys as a second argument: gvar.asbufferdict(g, keylist).


Version 4.5.3 - 2013-12-22
===========================

- Fixed bug in gvar._gvarcore that caused problems on win64 systems.

- GVar's __cinit__ has been changed to an __init__, which makes derivation
  from GVar possible. GVar also has new property: g.internaldata.
  This allows simplifies derivation from GVar --- see, for example,
  class WAvg in lsqfit._extras.py. Finally a cython declaration file,
  gvar.pxd, is installed for the benefit of other cython modules:
  cimport gvar gives the module access to the internal definitions of
  cython extension types GVar, svec and smat.

- lsqfit.wavg (weighted averages) now returns a variable
  of type WAvg which is a class derived from GVar (with all of
  its functionality) but with added attributes: chi2,
  dof, and Q which are the chi2, dof, Q from the wavg. In the past these were
  read off the function itself (eg, wavg.Q) but this is nonintuitive.
  Now ans = lsqfit.wavg(list_of_GVars) is a GVar with the extra
  attributes (ans.chi2, ans.dof, ans.Q). lsqfit.wavg still has attributes
  chi2, Q etc to help with legacy code. Also this is useful if the average
  is over a list of arrays or dictionaries (ie, a multidimensional random
  variable). In this case the individual GVars in the result have chi2s, etc
  as described above, while lsqfit.wavg has the chi2 for the entire set (ie,
  the sum of the chi2s for all the components).

Version 4.5.2 - 2013-09-26
==========================

- str(x) and repr(x) for GVar x both now return strings using the
  '2.31(10)' format rather than the older '2.31 +- 0.1'.
  The old format is still supported on input, but it will
  no longer appear in (default) printing. Use x.fmt(-1) to obtain the old
  format.

- Added gv.evalcorr(g) which calculates the correlation matrix of the
  GVars in g.

- gv.chi2 has a new option (fmt=True) that causes it to return a string
  (describing the chi**2) rather than the numerical value of chi**2.

- Operators > and < are now defined for gvar.GVars. This allows algorithms
  to order GVars, which is occasionally useful. The ordering is based upon
  the mean values. Operators >= and <= are still *not* defined, because of
  incompatibilities with == and !=, which look not just at mean values but
  also at all the dependencies. These incompatibilities suggest that one
  shouldn't define > and < either, except that there are times when it is
  quite useful to be able to order a numerical data type for algorithmic
  reasons. The setup here is a compromise (kludge?).

- Fixed very minor bug in lsqfit.nonlinear_fit.format().


Version 4.5.1 - 2013-08-13
==========================

- polishing/minor fixes for nonlinear_fit.simulated_fit_iter. Also now has
  a bootstrap option.

- copy.copy and copy.deepcopy now work with GVars.

- very minor fix to gvar.uncorrelated


Version 4.5 - 2013-07-31
========================

- nonlinear_fit.simulated_fit_iter generates fits of new simulated
  data that is generated randomly from the original fit data. This
  data is useful for testing fits and tuning parameters in them.
  Simulated data has the same covariance matrix as the original data but
  its mean values fluctuate around values given by the fitting
  function evaluated at user-specified parameter values p=pexact.
  The values in pexact are the "correct" values that should
  be obtained from a fit of the simulated data --- that is, the
  results of the fit to simulated data should agree with pexact
  to within errors. Knowing the correct answers for the fit
  parameters ahead of a fit allows for very realistic testing. See
  the documentation in the Tutorial section on Testing Fits with
  Simulated Data for more information.

- nonlinear_fit.format() now adds 1 to 5 stars at the end of any
  parameter line where the parameter and the prior differ by more
  than 1 to 5 (or more) standard deviations, respectively. Stars
  are also added when fit data is printed out where fit data
  and the fit differ by more than 1 standard deviation. These are
  meant to draw attention to potential problems.

- New function: gvar.chi2(g1, g2) computes the chi**2 of g1-g2, where
  g1 and g2 are (multi-dimensional) distributions. One of g1 or g2 can
  contain numbers instead of GVars (and/or can be missing entries
  contained in the other). Also gvar.chi2(diff) where diff = g1 - g2
  equals gvar.chi2(g1, g2).

- gvar.dataset.avg_data has new option specified by parameter noerror.
  Setting noerror=True causes avg_data to compute averages but not
  the errors in those averages.

- gvar.ranseed() called without an argument generates its own random
  seed to reinitialize the numpy random number generates. The seed is
  returned by the subroutine and can be used to recover the random
  number stream in later work. The seed is also stored in gvar.ranseed.seed.
  The idea is to use gv.ranseed() at the start of a code and print out
  gvar.ranseed.seed so that the seed can, if desired, be used to recreate
  the same random numbers in a later run. The key here is the 'if desired';
  usually you might not care to recreate a run unless something unusual
  happens.

- The tutorial in the documentation has a new section (at the end)
  with a pedagogical discussion of simple fit strategies.

Version 4.4.4 - 2013-07-07
==========================

- gvar.SVD sometimes complains that "SVD failed to converge". This is a
  numpy.linalg problem (that might be solved by *not* linking with atlas).
  Have introduced a back up routine (numpy.linalg.eigh) that is tried when
  this error is encountered.

- lsqfit.wavg now accepts a list of dictionaries (containing GVars or
  arrays of GVars), as well as lists of GVars or arrays of GVars.

- Modest optimization for gvar.evalcov. Small optimizaitons for gvar.svec
  and gvar.smat.

- Fixed bug in svec.add (where one or other svec is size=0 svec)

- Fixed very minor bug in gvar.gvar() (makes, eg, gvar(array(1.)) work).

Version 4.4.3 - 2013-04-06
==========================

- Improved syntax for @transform_p from lsqfit. The old syntax still works
  but the new syntax is simpler: 1) use @transform_p(priorkeys,0) instead
  of @transform(prior,0,'p'); and 2) fit.transformed_p is the same as
  fit.p but augmented with the exponentials of any log-normal terms, etc.

- Rules for initial values p0 in nonlinear_fit are more flexible: p0 can
  include keys that are not in prior (these will be ignored, unless prior
  is None). This makes it more likely that an old p0 will be useful for
  priming a new fit.

Version 4.4.2 - 2013-03-16
===========================
This is another minor upgrade.

- Evaluation of logGBF in nonlinear_fit was having problems (in one user's
  code, at least) with very large covariance matrices. This is now fixed.

Version 4.4.1 - 2013-03-14
==========================
This is a very minor upgrade.

- Set default svdcut=1e-15 instead of None in nonlinear_fit. This cut is
  very small and so usually has negligible impact in cases where an svdcut is
  unneeded. It protects against minor roundoff errors that arise relatively
  frequently, even in fairly simple problems. It also prevents problems from
  exact zero modes in the data or prior. One might argue that it would be
  useful to expose these last problems, rather than dealing with them quitely,
  but dealing with much more common minor roundoff errors seems more important.

- exp(fit.logGBF) is the probability (density) for generating
  the fit data from the input fit model, assuming Gaussian statistics.
  It used to be proportional to that probability; the
  proportionality factors are now included. This change will have no
  impact at all on almost all uses of logGBF. Change made more for the sake of
  clarity than utility.

- More documentation, including a tutorial section on chained fits and more
  discussion of svd cuts.

Version 4.4 --- 2013-02-13
==========================

- New function gvar.deriv(f, x) computes df/dx where f and x
  are gvar.GVars, and x is independent (ie, x has only one non-zero
  element in x.der). A ValueError exception is raised when x
  is dependent on other GVars. f can also be an array of GVars
  or a dictionary of GVars and/or arrays of GVars. GVars also
  have a method which computes the derivative: f.deriv(x).

- Small code improvements to lsqfit.transform_p.

Version 4.3.1 --- 2013-02-10
============================

- Slight refinements to the support for log-normal, etc
  priors. The decorator name is changed (but the old
  name is aliased to the new, to support legacy code
  (if there is any)).

Version 4.3 --- 2013-02-10
===========================

- Works with python3.3 (and numpy >= 1.17 which is necessary for 3.3).
  Fixed minor errors in gvar.BufferDict.__str__ and in some of the unittests
  that showed up with python3.3.

- Support for log-normal and "sqrt-normal" prior distributions for fit
  function parameters. The idea is to use parameters with names like
  "log(a)" instead of "a" in the prior, while expressing the fit
  function in terms of "a": so prior["log(a)"] is
  specified in the prior but not prior["a"], while the fit
  function uses parameter p["a"] but not p["log(a)"]. Parameter
  p["a"] has a log-normal distribution because prior["log(a)"] is
  a gaussian variable. See the section "Positive Parameters" in
  the overview section of the html documentation, for more
  information.

- gvar.dataset.Dataset changed to an OrderedDict from a dict. This mostly
  doesn't matter. Just about the only non-cosmetic effect concerns what
  happens when an svdcut is applied to the output of avg_data --- small
  differences arise when rows and columns of the covariance matrix are
  interchanged (roundoff error).

- Changed == and != for GVars to allow comparisons with non-GVar types; a GVar
  compares as not equal to a non-GVar unless its mean equals the
  non-GVar and its standard deviation is zero. Note that >, <, etc are
  not defined for GVars since GVars are not unambiguously ordered
  --- eg, a number drawn from the distribution 100(99) will be
  larger than one from 101(1) almost 50% of the time, even though
  100 < 101.

- Had too many pieces in the version number, so moved to 4.3. A
  third component, as in 4.3.1, will indicate bug fixes and minor
  features. There has been a lot added since 4.2 started (see 4.2.2).

Version 4.2.7.2 --- 2013-01-29
==============================
gvar.fmt_errbudget(...) has new parameter to specify column widths. This
allows for longer names for outputs and inputs.

Version 4.2.7.1 -- 2013-01-14
=============================
Adds a further tweak to the exception handling inside fit functions ---
slightly more robust than what is in 4.2.7.

Version 4.2.7 -- 2013-01-13
===========================
Another minor update:

- gvar.raniter and gvar.bootstrap_iter now work with single gvar.GVar's as
  arguments (in addition to the more useful cases of arrays and
  dictionaries). This makes them more consistent with the other utility
  functions.

- Python errors buried inside fit functions now result in slightly more
  intelligible error messages. Added two new unittests for such
  exception-handling.


Version 4.2.6 -- 2012-12-03
===========================
This is a minor update:

- Adds load (and loads) and dump (and dumps) methods to gvar.BufferDict to
  facilitate saving serialized BufferDicts in files (or strings) for later
  use. This is particularly useful when the BufferDict contains gvar.GVars
  since the correlations between the different GVars in the BufferDict are
  complicated to retain properly. These are implemented using pickle or,
  optionally, json. pickle already worked with BufferDicts. json was added
  because pickle is not compatible between python2 and python3. json files
  are also readable by non-python code (and by yaml). The json
  implementation has some limitations (around the types used for keys in
  the BufferDict, as well as types for the values) so pickle may be
  preferable except in situations where data must be moved from python2 to
  python3.

Version 4.2.4 -- 2012-08-18
===========================
This update is to fix a bug. Since version 4.2.2 lsqfit has been able to
deal correctly with statistical correlations between priors and the input
fit data. The code checks automatically for such correlations, and modifies
the definition of chi**2 appropriately if it finds correlations. There was
a bug in part of the code that checks for correlations, causing it to miss
certain situations. That bug is fixed in this update. Also

Other changes:

- Renamed gvar.orthogonal to gvar.uncorrelated, which is more intelligible
  (and also now has correct code).

- Fixed bug in gvar.GVar.partialvar (and therefore also
  gvar.fmt_errorbudget). The partial variance due to some GVar g should
  include the contributions from all other GVars that are statistically
  correlated with g. This previous code missed correlated but unreferenced
  variables that should have been included automatically.

- gvar.dataset.autocorr() is now done properly (with FFTs) and so can
  handle large datasets. It now computes autocorrelations for all
  intervals.

- lsqfit now issues deprecation warnings if the old classes GPrior,
  CGPrior, or LSQFit are used. These have been superseded in recent
  versions (by gvar.BufferDict and lsqfit.nonlinear_fit), and the old names
  have been attached to the new constructs, but the correspondence between
  old and new is only approximate --- hence the warning.

- Documentation improvements in the Tutorial.

Version 4.2.3 -- 2012-07-22
===========================
This version updates printing of GVars and of nonlinear_fits:

- Enhanced the formatting capabilities of GVar.fmt. If g is a GVar, then
  gvar.fmt() will create a string representation of g that shows the
  leading 2 digits of the error (used to be 1). The new code handles
  special cases much more effectively. For example very large or small
  numbers are represented using exponential notation (eg, 1.23(4)e+10 meaning
  1.23e+10 +- 4e+8). Also removed some bugs in the conversion from strings
  to GVars (eg, couldn't handle "-.2345(1)"). Added new unittests for fmt
  (in test_gvar.py).

- Changed the format of the fit report produced by
  nonlinear_fit.format(..). New format is more compact and more
  informative. In particular, indices for parameter arrays are included in
  the output to make finding a particular element easier. Also include
  errors on the fit values when data and fit are printed out. Output can be
  streamlined using new option pstyle='m'. (Setting pstyle='vv' gives
  output a lot like the old format.) Added unittests for format(..) (in
  test_lsqfit.py).

- Added new utility function gvar.fmt(g..) which formats every GVar in
  GVar/array/dictionary g (using x.fmt(..) for every GVar x in g).

- Scripts eg0.py ... eg5.py in doc/source now generate program output in
  files, with names like eg0.out and eg5b.out, that are read directly into
  the documentation. This simplifies the building of the documentation as
  changes are made to reporting functions (see above).

Version 4.2.2 -- 2012-06-07
===========================
This version involves significant internal change relative to the last
version, much of which will be invisible to most users. Significant pieces
of lsqfit and gvar were refactored for simplicity, with replacements for a
number of awkward constructions that reflected earlier but now obsolete
ideas about how the code would be used. A somewhat inconvenient change is
renaming the gdev module to gvar (for "gaussian variable"): every
instance of 'gdev' is now replaced by 'gvar', as is every 'GDev' by 'GVar'.
The old names were wrong and therefore misleading. (A tiny 'gdev.py' file
is included that aliases the new names with the old names, for use with old
code.) More usefully, the interfaces for many functions in lsqfit and
especially gvar were made more uniform: for example, almost any gvar
function that took an array of GVars as an argument can now also accept a
single GVar or a dictionary whose values are single GVars or arrays of
GVars. This is motivated by the overall design notion that multidimensional
distributions should be represented by collections of GVars: either as
arrays, or as dictionaries containing GVars and/or arrays of GVars, the
latter providing a much more flexible interface. These changes should make
the modules easier to learn and use, and certainly makes them easier to
maintain.

The bigger changes include:

- The names gdev and GDev are everywhere replaced by gvar and GVar (for
  "gaussian variable"). A new gdev.py module is included that aliases the
  new names to the old names, for use with old code. gdev.py is not
  installed with the rest of the code; if you need it (for old code)
  install it, for example, using "make install-gdev"; or copy it to the the
  directory containing the old code. Obviously, a better solution is to get
  rid of the old names.

- Correctly handles situations where priors are correlated with the fit
  data. Previously such correlations were ignored. This is the most
  significant change in functionality. It is a situation that arises rather
  rarely, but which is mishandled by older versions.

- Removed minor bug in lsqfit.wavg (used to ignore svdcut<0).

- Fit functions that depend only on the fit parameters (that is, have no
  dependence on an independent "x" variable) are now supported. This is
  signaled either by setting x=False in the fit data (data=(x,y)) or by
  leaving x out altogether (data=y) in nonlinear_fit.

- Rearranged gvar and lsqfit into packages instead of simple modules. This
  makes maintenance easier. It also reduces the number of names added to
  the module space.

- Relocated BufferDict into gvar. BufferDicts can still be constructed from
  dictionaries but no longer directly from arrays. This makes for a cleaner
  data type. BufferDicts are used internally in several of gvar's functions
  as the standard dictionary class (the standard array class is a numpy
  array). Unlike regular dictionaries, BufferDicts can be pickled even when
  filled with GVars; this is currently the only way to pickle GVars.

- Removed class GPrior from lsqfit. It isn't really needed any more since a
  dictionary works just as well. (GPrior is now an alias to
  gvar.BufferDict, which should allow older code to continue working,
  mostly.) Also removed classes BasePrior and NullPrior.

- svdcut and svdnum in nonlinear_fit still specify svd cuts for the fit
  data, but now can also specify svd cuts for the prior (no other easy way
  to do this now that GPriors are effectively gone). To specify a cut for
  the prior make svdcut and/or svdnum into 2-tuples, where the first entry
  is for the data and the second is for the priors.

- fit.svdcorrection is list with one or two elements. Either element can be
  a (1-d) vector or None. Can now be used directly as an input in
  fmt_errorbudget() (don't need/want to put [ ] around it).

- Merged class LSQFit and function nonlinear_fit from lsqfit into a new
  class called nonlinear_fit. nonlinear_fit is used as before, but is now
  actually initializing the class when it is fitting. Given standard usage,
  there was no reason to keep these two separate. (The old LSQFit class was
  originally meant to represent a fitter, but was mostly used to hold the
  results of a single fit; the new class nonlinear_fit class represents the
  result of a fit.)

- Redefined gvar.mean, gvar.sdev, gvar.var, gvar.evalcov, gvar.raniter, etc
  so that they all work with dictionaries as well as arrays. The
  dictionaries are converted to BufferDicts internally and results are
  returned as BufferDicts.

- The name of fmt_partialsdev is now changed to the more understandable
  fmt_errorbudget. Also it is part of module gvar, as well as being a
  method in nonlinear_fit objects. The name fmt_partialsdev is retained as
  an alias, to benefit older code.

- Allow arguments to GVar.partialvar and GVar.partialsdev to be None or
  single GVars or arrays/dictionaries of GVars. Arguments to
  gvar.fmt_errorbudget are also now allowed to be None, single GVars or
  lists of arrays/dictionaries of GVars. Previously each of these routines
  was more restrictive.

- Added a bootstrap_iter function to gvar to create bootstrap copies of
  collections of GVars (arrays or dictionaries).

- lsqfit's nonlinear_fit.bootstrap_iter does bootstrap fits on a list of
  bootstrap copies of the fit data. Now the list of bootstrapped data can
  be omitted and bootstrap copies are generated internally, from the means
  and covariance matrix of the data set. This is useful if the data has
  small errors (ie, is gaussian) which is often the case even if the fit
  parameters turn out to be non-gaussian (and therefore require
  bootstrapping).

- Created new options for gvar.gvar arguments: eg,
  gvar.gvar(["0(1)",(2,1)]) returns array [gvar(0,1),gvar(2,1)].

- Added new tools in gvar.dataset for handling random samples from
  distributions. These include functions avg_data(data),
  bootstrap_iter(data), and bin_data(data,binsize), as well as class
  Dataset for collecting random samples (in a dictionary). These additions
  are meant to supplant the old dataset.py module.

- Internal changes to how the data and covariance matrices are inverted
  could lead to small differences in results, due to roundoff error.

- nonlinear_fit.check_roundoff() now issues a warning, rather than an
  error, if large roundoff errors are suspected.

- svd analysis is handled by function gvar.svd which is now applied to a
  dictionary or array of GVars. It uses class gvar.SVD which is applied to
  a covariance matrix.

- nonlinear_fit.kappa no longer exists. It can be obtained using gvar.SVD.

- renamed nonlinear_fit.dump_parameters with nonlinear_fit.dump_pmean. Also
  added nonlinear_fit.dump_p and nonlinear_fit.load_parameters.

- Documentation streamlined. The Overview and Tutorial section was
  simplified a little, and has a new section on Troubleshooting.

- Speed is about the same except in cases where there are correlations
  between the priors and the fit data (where it is somewhat slower now,
  because it is doing the right thing).


# Created by G. Peter Lepage (Cornell University) on 2012-04-29.
# Copyright (c) 2008-2017 G. Peter Lepage.
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# any later version (see <http://www.gnu.org/licenses/>).
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
