CHANGELOG.org

Changelog

Unreleased

Added

Changed

Deprecated

Removed

Fixed

v0.3.17 - 2024-09-06

Added

• tetragono: Add neural network state based on sampling methods.
• tetraku: Add quantum chemistry model which reads Hamiltonian in openfermion format.
• tetragono: Add support for tracing repeated points when setting hamiltonian term. To use this feature, call state.hamiltonians.trace_repeated after all the hamiltonian term has been set. For those term which will not be traced, user could exchange the edge indices and points at the same time. For example a tensor with edges name as [O0, I0, O1, I1] on points (0,0,0) and (0,0,1), we can rename the tensor edge to [O1, I1, O0, I0] (it is rename, not transpose!), and apply it to points (0,0,1) and (0,0,0), and the system does not change. But for terms which may be traced, we cannot exchange them. The order of edge indices hint the rule how to trace if points duplicated.
• tetragono: Add support to combine hamiltonian terms on the same group of points with different orders by sort_points, which make the program faster for simple update and exact update.
• tetragono: Add support to check whether the hamiltonians is hermitian by check_hermite(threshold), which requires sort_points invoked first.

Removed

• tetragono: Remove SIGINT handle in gradient descent.

Fixed

• tetragono: Fix the import issue in tetragono.
• tetragono: Fix missing signal SIGUSR1 in windows platform.

v0.3.16 - 2024-03-20

Fixed

• tetragono: Fix the missing conjugate for the energy during calculation of natural gradient.
• tetragono: Fix the compatibility for old version python. Some previous codes uses syntax introduced in python 3.11.

v0.3.15 - 2024-02-28

Added

• TAT.py: Add const_blocks for tensor to get the const block which will not check the ownership.
• tetragono: Add __contains__ and __len__ for hamiltonians handle of abstract state.
• bridge: Add support for bridging from TNSP new format.
• tetragono: Stop saving configuration for ergodic sampling method during gradient descent.
• tetragono: Add function to convert sampling lattice back to simple update lattice.
• tetragono: Add support for simple update for long range two body interaction.

Changed

• TAT.hpp: Rename the aliases of the symmetry types, the details are shown in the next item.
• TAT.py: Rename the names of the symmetry types, since the previous naming is not consistent:

  old name	new name
  No	No
  Z2	BoseZ2
  U1	BoseU1
  Fermi	FermiU1
  FemriZ2	FermiU1BoseZ2
  FermiU1	FermiU1BoseU1
  Parity	FermiZ2
  FermiFermi	FermiU1FermiU1

• TAT.py: Use scikit-build-core as build backend. MAKEFLAGS and CMAKEFLAGS are droped. Use scikit-build-core environment variable SKBUILD_CMAKE_ARGS to pass arguments to cmake command line when building PyTAT wheel.
• tetragono: Use Lanczos algorithm instead of previous power iteration to optimize exact state. update function for exact state does not have the second parameter approximate_energy any more, and ex_update in tetragono shell does not have the second parameter either.
• TAT.hpp: Change the default exponential iteration step to 8 from 2, since in practice, it is difficult for step=2 to convege to a usable result.

Removed

• tetragono: Remove merging Hamiltonians before simple update automatically.

Fixed

• TAT.py: Fix a bug in loading Symmetry or Edge.
• bridge: Fix several bugs in tnsp bridge, because of the TAT interface changes a lot since the first release of tnsp bridge.
• TAT.py Fix wrong arguments number for function __deepcopy__ of tensor.
• TAT.hpp: Fix a bug when reading tensors from text data with ANSI control char mixed.

v0.3.12 - 2023-11-29

Added

• TAT.py: edges[int] and edge_by_name(str) are the new interface for indexing edges in tensor, the previous interface has been deprecated.
• TAT.py: Add copy constructor for Symmetry, Edge and Tensor.

Changed

• TAT.hpp: The names of all inplace functions are renamed with a suffix _.
• TAT.py: The names of all inplace functions are renamed with a suffix _.

Deprecated

• TAT.py: edges(str) and edges(int) is deprecated for tensor.

Removed

• tetraux: Package tetraux is removed, because it is useless since ansatz product state removed.
• tetragono: gm_bin_run and gm_alpha are removed. gm_bin_run is removed since it only works for sweep sampling, but there is a better way to do so: sampling multiple Markov Chain in the same process. gm_alpha is removed since user should set attribute of the state manually instead.

Fixed

• TAT.hpp: Fix type mutability information in declaration of LAPACK function ?gesvd.

v0.3.11 - 2023-11-22

Added

• tetragono: Add gm_conf_eq to equilibrium the configuration of the sampling lattice.
• TAT.py: Add TAT.parity, TAT.arrow to convert +1 to False and -1 to True.

Changed

• tetragono: The physics edge index should be non-negative now.
• tetragono: Use MPI IO to speed up the performance of configurations. Previously, every process will read all configuration and choose a single configuration when reading, and allgather all configuration and write to file when writing. Now, the program uses MPI IO to write file parallelly, and read the corresponding configuration directly without read all configurations.
• tetragono: Yield measurement result before save state and configuration in gradient descent of sampling lattice. Since user may want to update attributes out of yield.
• tetragono: Allow user to set hamiltonian term on the same positions multiple times, which allows user to use state.hamiltonians[positions]+=tensor. Previously, the program will throw error saying hamiltonian term has already been set, but recently, some models need adding term on the same positions.

Removed

• tetragono: Remove ansatz product state support, which may be added again in the future, but it require many updates.

Fixed

• tetragono: Fix bug when total sampling number less than process number in sampling gradient.

v0.3.9 - 2023-09-07

Changed

• tetragono: The interface of save_result allow only (state, result, whole_result) now.
• TAT.hpp: Change the view of sign problem in conjugate, there is only one conjugate, but the trivial metric may not be positive definite, so in gradient descent, this negative metric term need to be considered, the optional paramter trivial_metric is used to apply the trivial metric when conjugating.
• TAT.py: Update the new interface of conjugate.
• TAT.py: Use customized string stream in TAT.py to avoid a duplicated copy of string.
• TAT.hpp: Check the larger tensor to try to fit the common edges in contract, if not proper, check the other one. The previous behavior always check tensor 1 firstly and check tensor 2 secondly.
• scalapack.py: Change the loader API. The previous way is from PyScalapack import Scalapack and using Scalapack("libscalapack.so") to get the scalapack object, Now user should use import PyScalapack and PyScalapack("libscalapack.so") directly.

Removed

• TAT.hpp: Remove deprecated tensor member function get_rank, find_rank_from_name and get_rank_from_name.

Fixed

• TAT.hpp: Fix a bug about parity sign when merge_map is empty.
• TAT.hpp: Fix invalid access to a empty vector when splitting the first edge into zero edges.
• TAT.hpp: Fix out of range pointer error when try to split a trivial edge {{Symmetry(), 1}} to no edge, or merge no edge into a trivial edge.
• TAT.hpp: Fix an out of range bug in identity.hpp. It created a pod vector with memory reserved and use the item directly previously, without resize, which would not cause critical issue in fact. Emplace the item when using now.

v0.3.7 - 2023-08-15

Added

• tetragono: Add support to calculate deviation of the observer set as a whole, which is useful since many observables are the summation of several local observables.
• tetragono: Add arguments support for gm_conf_create and ap_conf_create in tetragono shell.
• tetragono: Add a member dict called attribute to AbstractState, which may contain customed data used by user.

Changed

• tetragono: save_result accept another parameter, which is the total result of the whole observers set. Its signature is state, result, whole_result and step now.
• tetraku: Change the defnition of x/y/z link in kitaev model. The bond inside one tensor is z link now.
• tetragono: gm_run(_g) will now return/yield all the measurement result instead of energy only.

Deprecated

• tetragono: The previous signature of save_result: state, result and step is deprecated.

Removed

• tetragono: Remove the deprecated function natural_gradient and the deprecated paramter metric_inverse_epsilon for gm_run.

v0.3.6 - 2023-05-26

Added

• TAT.hpp: Add clear_fermi_symmetry for tensor with fermionic symmetry, which convert the tensor to the simplest fermionic tensor, tensor with parity symmetry.
• TAT.py: Add python binding for clear_fermi_symmetry.
• tetragono: Check whether the state is None when dump the state to file to avoid overwriting file by mistake.
• tetragono: Add gm_clear_symmetry for sampling lattice.
• tetragono: Add gm_run_g and ap_run_g for middle level API. These are generators, which yield the energy of the state before gradient descent after every gradient step.
• TAT.hpp: Cut is used directly for the dimension cut in svd. Cut will detect the type of the input to determine whether it is relative cut or remain cut, if nothing provided, the behavior of previous NoCut will be used. And user could also set remain cut and relative cut together.

Changed

• tetragono: new_dimension in simple update will be intepreted as the threshold of singular value when it is float.
• TAT.hpp, TAT.py: clear_symmetry will return NoSymmetry tensor or ParityTensor based on whether the original tensor is fermionic. The previous behavior would give NoSymmetry tensor only, which is dangerous for fermionic tensor, is renamed to clear_bose_symmetry.

Deprecated

• TAT.hpp: NoCut, RemainCut, RelativeCut is deprecated, please use Cut directly.

Fixed

• tetragono: Fix a bug in merging physical edges during simple update. When the hamiltonian on single site with a large dimension is operated, the dead loop would occur previously.

v0.3.5 - 2023-04-01

Added

• tetragono: Add common used tensors for parity symmetry and some related tensor in Hubbard model with parity symmetry.
• scalapack.py: Add documents for scalapack.py.
• tetraku: Add tV model.
• tetragono Add numpy_hamiltonian in tetragono shell to export the Hamiltonian of a model to a file in numpy array format.

Changed

• tetraku: Density matrix of Heisenberg/tJ/Hubbard model are renamed to gibbs_* from density_matrix_*.
• tetragono: Simple update will check to avoid merge physical edges into a single edge with very large dimension. It will only merge edges if the result edge has the dimension less equal to the threshold, which is 6 by default.
• tetragono: Use better method to estimate the variance of measurement results observed by the sampling method.
• tetragono: Tetragono shell will not refresh the corresponding configuration when creating or loading ap or gm state. Since sometimes user may want to share the configuration between states.
• tetragono: Use pseudo inverse directly in conjugate gradient method.
• TAT.py: Use mt19937(64 bit) as the random generator instead of c++ default random generator which may be different on different platforms.

Removed

• tetragono: The mirror direct sampling is removed, since it is useless according to the test result.

Fixed

• tetragono: Fix bug in min-SR method when the model is complex, A conjugate operator was missing.
• tetragono: Avoid nan in pseudo inverse of min-SR by checking whether the eigenvalue is zero first.
• tetragono: Fix the missing imaginary part of energy when calculating gradient.
• tetraku: Fix the wrong Hamiltonian for Gibbs state of Fermionic system.
• tetraku: Fix the wrong sign of chemical potential for density matrix of tJ model.
• TAT.py: Fix typo in the class name of parity symmetry.

v0.3.3 - 2022-03-09

Added

• tetragono: Add mirror direct sampling which is only used for Gibbs state on square lattice, and maintains the symmetry of the Gibbs state.
• TAT.py: Add dtype and btype class member for Tensor object, which makes it easier to communicate with numpy.
• scalapack.py: Add a python wrapper for scalapack.

Changed

• tetragono: Use the PyScalapack to speed up min-SR method. User need to specify the path of scalapack dynamic link libraries by parameter scalapack_libraries for gm_run when natural_gradient_by_direct_pseudo_inverse enabled.
• TAT.py: Change the module alias name convension, float and complex without bytes specified would be considered as double precision now.

Fixed

• tetragono: Fix a bug when try to save a file with directory name. The previous program only allows to save file into the current directory.

v0.3.2 - 2022-02-28

Added

• tetragono: Add natural_gradient_by_direct_pseudo_inverse to calculate natural gradient for sampling lattice. And add parameter use_natural_gradient_by_direct_pseudo_inverse (default is False), natural_gradient_r_pinv and natural_gradient_a_pinv for high/mid-level API in gm_run to use direct pseudo inverse to calculate natural gradient.
• tetraku: Add Hamiltonian to another part of density matrix for density matrix of Heisenberg/Hubbard/tJ model. It is to ensure the result density is unitary despite of errors introduced by contract and approximation. This is controled by a new parameter side which is either 1 or 2, default is 1, which has the same behavior with before.

Changed

• TAT.py: sqrt will calculate the square root of absoluate value in tensor elementwisely, instead of square root of value itself and return nan for negative number.
• tetraux: Move Configuration for ansatz product state to an individual package named tetraux from TAT.py, which is not related to the tensor itself.

Deprecated

• tetragono: natural_gradient for observer object is deprecated, users should specify the method to calculate natural gradient explicitly, which are natural_gradient_by_direct_pseudo_inverse and natural_gradient_by_conjugate_gradient.

v0.3.1 - 2023-02-21

Added

• TAT.py: Add binding for functions of Edge introduced in v0.3.0 such as point_by_index.

Changed

• TAT.py: Update the function arguments names to keep the same with those in c++ side.
• TAT.py: Remove navigator of TAT.py to get tensor type directly, please use module alias instead. For example, previous code such as `TAT(“No”, np.float64)` should be updated to `TAT.Normal.float64.Tensor`.

Removed

• TAT.py: Remove optional FastName binding, which is useless in python side in fact.

v0.3.0 - 2022-11-15

Added

• tetragono: Tetragono will print backtrace of the current process when receiving SIGUSR1.
• tetragono: Add squash support for sampling lattice.

Changed

• TAT.hpp: Use the multidimension span to record blocks in tensor, instead of the previous map data structure, some related API is also updated. Detail update is followed:
  • About data
    • Tensor blocks is stored in a new order other than old version, the previous use a map from symmetry list to data block, which follows the lexicographical order of symmetry list. The new order follows the lexicographical order of the symmetry position list for a data block. Inside the data structure, the blocks are stored in a simple and raw tensor like structure called multidimension_span.
    • Because of the block order update, use random number to fill a tensor will return a different one other than the previous version even with the same random seed.
    • The edge is now assumed stable, That is to say the edge will not lose any segment during operations. In the previous version, the edge segment will be erased if no block in the tensor using that segment.
  • About edge API
    • The type edge_segment_t is renamed to edge_segments_t because it is really several segments, not only one segment.
    • Some old function was renamed, such as get_point_from_index to point_by_index. The old name is deprecated and will be removed later.
    • Drop the support for reorder segments.
    • Use edge.segment() to obtain the real segment for an edge, instead of the original way to access member edge.segment directly.
  • About tensor API
    • Some old function was renamed, such as get_rank_from_name to rank_by_name. The old name is deprecated and will be removed later.
    • Use tensor.names() to obtain the tensor edge names, instead of the original way to access member tensor.names directly.
    • Because the edge is stable now, scalar operations on two tensor with segment and block mising is not allowed now.
• tetragono: Update line search strategy, remove line_search_error_threshold, add line_search_parameter in ap_run and gm_run. line_search_parameter multiplied step_size obtained by line search will be the real step size used to update the state.

Deprecated

Removed

• tetragono: gm_data_load is removed, please use gm_hamiltonian to replace the hamiltonian instead.
• wrapper: wrapper_TAT is removed.

Fixed

• tetragono: Fix the wrong error message when trying to import module used by ex_create, ap_ansatz_mul and so on.

v0.2.23 - 2022-10-09

Added

• tetragono: Add ap_hamiltonian to replace the hamiltonian of the ansatz product state in tetragono shell.
• tetragono: Add multichain_number for ap_run, which will run multiple chains inside the same MPI process.
• wrapper: Add python package wrapper_TAT to provide a wrapper over torch to provide similar interface as TAT.py.
• tetragono: Add observe_max_batch_size option for ap_run, which will set the max limit of batch size when calculating wss.

Deprecated

• tetragono: gm_data_load is deprecated, it will be removed in the future, please use gm_hamiltonian to replace the hamiltonian instead.

Removed

• tetragono: save_state_interval option for gm_run and ap_run is removed. The state will be saved for every step.

v0.2.22 - 2022-09-02

Added

• tetragono: Add save_configuration_file option for gm_run and ap_run in tetragono shell, which saves sampling configurations during gradient descent.
• tetragono: Add list as interface for rename_io in tetragono.common_tensor.tensor_toolkit. Original argument such as {0: a, 1: b, 2: c} can be written as [a, b, c].

Deprecated

• tetragono: save_state_interval option for gm_run and ap_run is deprecated. The state will be saved for every step if save_state_file was not set in the future.

Removed

• tetragono: The original function name create to create lattice is removed, which was deprecated in v0.2.18. The new function name to create lattice is abstract_lattice.
• tetragono: _owner of Configuration for sampling lattice and ansatz product state is removed, use owner instead.

Fixed

• TAT.hpp: Fix an internal compiler error for some old compiler, caused by the feature: fusing edges during tracing.

v0.2.20 - 2022-08-02

Added

• tetraku: Add models data and ansatzes data into an individual package named tetraku.
• tetragono: Configuration use owner to get the owner sampling lattice object of this configuration object, instead of the previous _owner.
• TAT.hpp: Add fusing edges support when trace a tensor, to keep the consistency with function contract.
• TAT.py: Add fusing edges argument binding for function trace of the tensor.

Changed

• tetragono: Rename multiple product state to ansatz product state, to avoid the ambiguous abbreivation. Rename all mp_xxx to ap_xxx in tetragono shell.
• TAT.hpp: Two new internal names used by user customed name type are added: Trace_4 and Trace_5. For the simple internal name usage, two new default internal names are added: Default_3 and Default_5.

Deprecated

• tetragono: _owner of Configuration for sampling lattice is deprecated, use owner instead.

Fixed

• TAT.hpp: Fix a bug in windows platform when copying an edge with fermi symmetry.

v0.2.19 - 2022-07-13

Added

• tetragono: Add a new command gm_hamiltonian to replace the Hamiltonian of the existent sampling lattice.
• tetragono: Add conjugate_gradient_method_error option for gm_run and mp_run in teragono shell. The conjugate gradient will stop if conjugate_gradient_method_step reached OR conjugate_gradient_method_error reached. Set conjugate_gradient_method_error to 0.0 to skip error checking or set conjugate_gradient_method_step to -1 to skip step checking.

Changed

• lazy: Using a manual stack to run the recursion now, to avoid the recursion depth limit.

Fixed

• tetragono: Fix a problem when calling gm_data_load in tetragono shell.
• tetragono: Fix a bug in calculating the natural gradient of a complex tensor network state.
• tetragono: Fix a bug in calculating the expect and the deviation in the ergodic sampling with subspace restricted.

v0.2.18 - 2022-06-28

Added

• tetragono: Add compatibility support for python3.8.
• tetragono: Add ex_create command for tetragono shell, which calls abstract_state to create an exact_lattice.
• tetragono: Add the multiple product state, which is the supertype of the string bond state and the convolutional neural network state, and add its high-level API in tetragono.shell.

Changed

• tetragono: The function name to create a lattice changes from create to abstract_lattice. And it is recommended to split it into two parts: abstract_state to create an abstract state and abstract_lattice to convert the abstract state into an abstract lattice.
• tetragono: The epsilon to avoid the singularity of the metric in the natural gradient method is relative now.

Deprecated

• tetragono: The original function name create to create lattice is deprecated.

v0.2.17 - 2022-06-12

Added

• tetragono: Add support for changing the hamiltonian of an existent sampling lattice. To do it, create a lattice with target hamiltonians and then call gm_data_load(file_name) to replace tensors with data stored in file file_name. For low-level API, lattice_1._lattice = lattice_2._lattice could replace tensor data directly.
• TAT.py: Add edge fusing support when contracting two tensors.
• TAT.hpp: Add support for being compiled by MSVC or Intel ICC.

Changed

• tetragono: Avoid data files being destroyed if a file system error appears when saving, such as the size limit of quota reached.

Removed

• tetragono: Submodule common_variable is removed, please use common_tensor or common_toolkit instead.

Fixed

• TAT.hpp: Fix selecting the wrong constructor of Edge when using two vector iterators to construct Edge.
• TAT.hpp: Fix bug when tensor SVD with relative cut cutting nothing.

v0.2.13 - 2022-04-26

Added

• tetragono: Allow passing function directly instead of module name string of various interface of mid-level API, including restrict for restricting subspace, measurement for measuring values, initial_configuration for initializing sampling configurations, and hopping_hamiltonians for fake hamiltonians used by sweep sampling.
• tetragono: Add load_configuration to load configuration from a saved file to a Configuration object for low-level API.
• tetragono: Add classical term of energy for sampling lattice. To do it, call observer.set_classical_energy(classical_energy) to set a function with configuration as input, and returns a float as classical energy. For mid-level API, pass a module containing function classical_energy or the function itself as parameter classical_energy to gm_run or gradient_descent.

Changed

• tetragono: Using hopping_hamiltonians instead of original hamiltonians for the name of function which generates fake hamiltonians used by sweep sampling.
• tetragono: Updating interface for generating initial sampling configuration(initial_configuration). Previously, state and Dc are given, it is needed to create Configuration manually. Currently, the input function will get initiated Configuration, and set configuration on it directly.
• tetragono: Configuration information is maintained by three mid-level API now: gm_conf_create, gm_conf_load, and gm_conf_dump. Call gm_conf_load(file_name) to load configuration from file. Call gm_conf_dump(file_name) to dump configuration to file after gm_run. Call gm_conf_create(initial_configuration) to using function initial_configuration to initialize configuration.
• tetragono: Rename parameter name sweep_initial_configuration to sampling_configurations. Generally, this parameter is not used directly by mid-level API, and it is needed to use gm_conf_create/load/dump to manipulate it. In the low-level API environment, load_configuration may be used to create configuration passed to sampling_configurations.