v0.3.3

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

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

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

Added

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

Changed

• TAT: 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

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

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 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

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

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

TAT 0.2.18

v0.2.17

TAT 0.2.17