Version 0.15.4

This release should have no functional changes for users from previous releases. Some small modifications have been made for C++ 20 support. In addition significant additions to the package were made to enable unit testing on GitHub for developers.

Version 0.15.3

Reorder looping structures in the Amplitude and AmplitudeManager classes to avoid CPU cache misses when fetching from memory. This version should be functionally identical to previous versions but provides some significant peformance improvements for some CPU fits.

Version 0.15.2

A number of bug fixes have been made in this version. One of the most
significant issues is to modify the likelihood() method of the AmpToolsInterface
class to fetch the likelihood using the same mechanism as is used by MINUIT.
This guarantees that the likelihood includes extra contributions beyond
those provided by the LikelihoodCalculator (e.g., Gaussian bounds or
user defined likelihoood contributions). A few other memory mangement issues
were fixed.

Version 0.15.1

This version fixes a bug in the splash screen logic. It also provides the user an option to trigger recalculation of user data by invalidating the calculations that are stored in the AmpVecs cache. This is necessary in the case that one wants to scan parameters and user data depends on
the value of some fixed parameter. Finally, there is an option to turn off the warning about intensity errors in FitResults.

Version 0.15.0

This new version now supports the ability for users to write a customized
contribution to the calculation of -2 ln( L ). (Thanks to Nils Huesken
for developing this capability.) Users can create a class that inherits
from Neg2LnLikContrib in much the same way that user Amplitude classes
are created. This contribution is then initalized with a neg2LnLik
keyword in the config file and parameters can be included. This may be
useful to use external data, e.g., published data with errors, to
provide additional constraints on parameters in the context of the
main unbinned likelihood fit. An example is included with the Constraint
class in the Dalitz tutorial.

Additionally, a script is now provided in the ROOT directory that will
load a subset of AmpTools modules into ROOT. Of particular use are the
FitResults and ConfigurationInfo objects.

Finally, the only other significant functional change is that when
parameters are reinitialized or randomized (using the AmpToolsInterface)
an underlying call to MNRSET in Minuit is made. This forces MINUIT to
redetermine step sizes in minimization. This may lead to better outcomes
when repeated fits are performed with randomized parameters.

Version 0.14.5

This release has some small corrections to the Dalitz tutorial to make it more consistent with documentation. A new reporting class for formatting screen output is introduced. This makes running large-scale MPI jobs posssible as much of the output on follower nodes is suppressed. Finally use of auto pointers is removed in some places to help with backwards compatibility for older compilers. There are no functional changes to core fitting methods.

Version 0.14.4

This is another bug fix that corrects scaling of MC histograms in the
PlotGenerator, which was affected by rescaling of the likelihood that was
introduced in v0.14

Version 0.14.3

This release contains a bug fix that leads to bogus results when computing the intensity using the FitResults class. Fits may appear to work fine, but due to a read of unitialized memory in the normalization integral matrix one can obtain unphysical and nan results for intensity computations.

Version 0.14.2

This re-enables new likelihood scaling. Consistent likelihood results now are report for MPI and non-MPI jobs. No other functionality changes have been made.

Version 0.14.1

This release turns off by deafult the new likelihood scaling. That is,
in the context of the notes below, it sets LEGACY_LN_LIK=1 by default.
The scaling is functional, but as implemented results in different
values for the likelihood at minimum for MPI vs. non-MPI jobs due to
how the scale factors are computed. While the fit results are fine,
this variation is bothersome for production use and needs to be fixed
prior to making this the default.