Tags: krivenko/som
Tags
Major release 2.0.0
Major changes and new features
==============================
- Complete port to TRIQS 3.1 and Python 3.
- Implementation of the Stochastic Optimization with Consistent Constraints
(SOCC) proposed by Goulko et al in Phys. Rev. B 95, 014102 (2017).
It includes three major pieces of functionality.
* The Consistent Constraints update in the Markov chain used to accumulate
particular solutions;
* The Consistent Constraints protocol for constructing final solutions out of
particular solutions;
* The solution quality assessment technique implemented in a new Python module
`som.spectral_stats`.
- For consistency with MaxEnt and other stochastic continuation methods, the
objective function of the optimization problem has been changed to the
"goodness of fit" $\chi^2$-functional.
- Adoption of the $\chi^2$-functional has made it possible to support
user-supplied covariance matrices of input data as an alternative to simple
estimated error bars (credits to @snirgaz for proposing this feature).
- A new family of integral kernels for symmetric fermionic Green's functions has
been introduced. The corresponding observable is called `FermionGfSymm`.
- The `BosonAutoCorr` kernels have been changed to more closely reproduce
results of the `BosonCorr` kernels for the same input data. Both kernel
families are defined on the whole energy axis and expect the same spectrum
normalization constants now (before one had to divide the constants by 2 for
`BosonAutoCorr`).
- Projection of an observable onto a real frequency mesh can now be
performed using binning (enabled by default). In this mode the projected
observable is integrated over bins centered around points of the mesh.
- Further MPI parallelization.
- Massively reworked online documentation.
Python API changes
==================
- Following a convention change for TRIQS applications, the Python package of
SOM has been renamed from `pytriqs.applications.analytical_continuation.som`
to a laconic `som`.
- Functionality of the `run()` method of `SomCore` has been split among a
few new methods,
* `accumulate()` -- accumulate particular solutions;
* `adjust_f()` -- adjust the number of global updates `F`;
* `compute_final_solution()` -- construct the final solution using the
standard SOM protocol;
* `compute_final_solution_cc()` -- construct the final solution using
Goulko's SOCC protocol.
One may still call the deprecated `run()`, which is equivalent to calling
`accumulate()` + `compute_final_solution()`.
- In recent versions of TRIQS it became impossible to use the `<<` syntax to
fill GF containers from user-defined Python objects. Furthermore,
high-frequency tail data was separated from the GF containers. As a result,
that syntax had to be abandoned in favor of a few free functions.
* `fill_refreq()` -- fill a real-frequency observable from a computed SOM
solution;
* `compute_tail()` -- compute high-frequency tail coefficients from a
computed SOM solution;
* `reconstruct()` -- reconstruct input from a computed SOM solution.
- It is now possible to resume accumulation of particular solutions by calling
`SomCore.accumulate()` multiple times, and to discard all accumulated
solutions by calling `SomCore.clear()`.
- A handful of new properties and accessor methods have been added to `SomCore`.
- The `rectangle` and `configuration` C++ objects are now exposed as Python
classes `Rectangle` and `Configuration`. `Configuration` objects can be saved
to/loaded from HDF5 archives.
- Updated signature of `som.count_good_solutions()` to take both
`good_chi_abs` and `good_chi_rel` (thresholds on $\chi$ and
$\chi/\chi_\mathrm{min}$ for a solution to be considered good).
- A new utility function `som.estimate_boson_corr_spectrum_norms()` has been
added. Given a correlator of boson-like operators $\chi$ defined on any
supported mesh, it returns a list of spectrum normalization constants
$\mathcal{N} = \pi \chi(i\Omega = 0)$.
Build system and developer tools
================================
- Minimum required CMake version has been bumped to 3.12.4.
- Structure of the project has been adjusted to follow conventions established
by the app4triqs application template.
- A `Dockerfile` has been added.
- Files `som.modulefile` and `somvars.sh` are generated and installed as part
of the build process.
- New benchmarks: `all_kernels`, `binning`, `consistent_constraints`,
`bosonautocorr` and `fermiongfsymm`.
- The `chi` benchmark has been removed as it depended on the private
`triqs_ctseg` code.
- Support for C++ static analysis tools `clang-tidy` and `cppcheck` has been
added.
- A CMake option has been added to link `libsom` and unit tests to Clang
sanitizers (`AddressSanitizer` and `UndefinedBehaviorSanitizer`).
- C++/Python coding style is enforced with `clang-format` and `flake8`.
Release 1.2 This is the last SOM release compatible with TRIQS 1.4.x. - Improvements and small fixes in documentation. - Added a new histogram post-processing function, `count_good_solutions()`. - Fixed a bug in `update_glue_shift` elementary update. - New benchmark `all_kernels` and fixes in the `chi` benchmark. - Added Travis CI config for continuous testing and documentation deployment. - Minor code improvements.