1 files changed, 123 insertions, 0 deletions
diff --git a/CHANGELOG.rst b/CHANGELOG.rst
new file mode 100644
index 000000000000..d95715b8ae4d
--- /dev/null
+++ b/CHANGELOG.rst
@@ -0,0 +1,123 @@
+Changes from version 4.1 to 4.2
+===============================
+
+This update mostly involves some improvements to the new patch-based
+computations, in response to feedback from users, who have been trying it out
+in a variety of use cases.  Especially note the new optional ``func`` argument
+to `BinnedCorr2.estimate_cov` and `estimate_multi_cov`, which lets you
+calculate covariances of arbitrary functions of a correlation object or list
+of correlation objects.  See `Covariance of Derived Quantities` for details.
+
+Another big improvement is the ability to correctly perform three-point
+cross-correlations. I still don't have heterogeneous cross correlations like
+NGG or NNG, but now NNN, GGG, and KKK can be done correctly for 2 or 3
+different catalogs.  This update does technically involve an API change,
+since the old code tried to do something when given 3 catalogs, but it
+wasn't really correct, and (consequently) I don't think anyone was actually
+trying to use that feature.
+
+See the listing below for the complete list of new features and changes.
+`Relevant PRs and Issues,
+<https://github.com/rmjarvis/TreeCorr/milestone/7?closed=1>`_
+whose issue numbers are listed below for the relevant items.
+
+.. note::
+
+    This will be the last (non-bug-fix) TreeCorr release to support Python 2.7.
+    It is becoming harder to continue to support this platform now that it is
+    officially sunsetted, so we encourage all users to switch their code to
+    Python 3.x ASAP.  TreeCorr is currently compatible with Python versions
+    2.7, 3.5, 3.6, 3.7, 3.8, 3.9.  We will probably also drop 3.5 in the next
+    release as well, since that is also retired at this point.
+
+API Changes
+-----------
+
+- Changed the 3pt process function, when given 3 arguments, to accumulate all
+  triangles where points from the three input catalogs can fall into any
+  of the three triangle corners.  If you need the old behavior of keeping
+  track of which catalog goes into which triangle corner, the new
+  ???CrossCorrelation classes do so both more efficiently and more
+  accurately than the previous code. (#115)
+- Changed the `NNNCorrelation.calculateZeta` function to only take
+  RDD and DRR parameters for the cross terms (if desired) rather than all
+  six (DDR, DRD, RDD, DRR, RDR, RRD).  The new cross-correlation behavior of
+  the `NNNCorrelation.process` function now efficiently calculates in RDD what
+  used to be calculated in three calls for DDR, DRD, and RDD.  Likewise the
+  new DRR calculates what used to require DRR, RDR, and RRD. (#122)
+
+Performance improvements
+------------------------
+
+- Only show at most 10 rows with NaNs in cases where there are lots of such
+  rows. (#111)
+- No longer remakes patches and writes them to disk if they are already present
+  on disk.  If you want to force a rewrite for any reason, you can explicitly
+  call `Catalog.write_patches`. (#116)
+- Computing the data/random cross correlations for 3pt are now much faster,
+  since you only need one call for RDD and one for DRR, not all the 6 different
+  permuations. (Specifically, DDR, DRD, RDR, RRD are no longer needed.) (#122)
+
+New features
+------------
+
+- Implement patch-based calculations for 3-point correlations.  This enables
+  things like jackknife correlations, low-memory calculations, and MPI execution
+  when doing NNN, KKK, or GGG correlations. (#97, #125)
+- Allow min_rpar and max_rpar for Euclidean metric.  Also Periodic, although
+  I don't know if that's useful. (#101)
+- Added ability to read from hdf5 catalogs.  (#106)
+- Added ability to use named columns in ASCII files if the file has column
+  names. (#108)
+- Added optional initialize=False and finalize=False options to process
+  functions. (#109)
+- Added ability to do 3 point cross-correlations properly, rather than the not
+  exactly correct version that had been implemented.  Not all combinations are
+  implemented.  So far just GGG, KKK, and NNN. (#115)
+- Added ability to read from parquet catalogs.  (#117)
+- Added optional func parameter to estimate_cov and estimate_multi_cov. (#118)
+- Added ability to write output files as hdf5.  (#122)
+
+Bug fixes
+---------
+
+- Fixed a couple problems that could arise in patch-based calculations when
+  one or more of the patches have no objects. (#123)
+
+
+Changes from version 4.2.0 to 4.2.1
+===================================
+
+- Fixed a bug when running NN correlations with MPI.  Thanks to Joe Zuntz for
+  the bug report. (#127)
+
+Changes from version 4.2.1 to 4.2.2
+===================================
+
+- Fixed a build problem with gcc versions 10 and 11.  Thanks to Johannes Lange
+  for reporting the problem.
+
+Changes from version 4.2.2 to 4.2.3
+===================================
+
+- Fixed a bug in the Map^3 calculation when the aperture radii are different.
+  Thanks to Laila Linke for both finding the problem and providing the fix. (#128)
+
+Changes from version 4.2.3 to 4.2.4
+===================================
+
+- Fixed a bug when running the Rlens metric with patches. Thanks to Marina Ricci
+  for the bug report. (#131)
+
+Changes from version 4.2.4 to 4.2.5
+===================================
+
+- Fixed bugs in the shot noise variance estimate for GGG and the propagation of
+  the error to Map3.  Thanks to Lucas Secco for the bug report. (#133)
+
+Changes from version 4.2.5 to 4.2.6
+===================================
+
+- Fixed a numerical bug that could suppress the signal in GG correlations in
+  spherical coordinates if there are multiple objects at the same position.
+  Thanks to Sandra Unruh for the bug report. (#134)