Skip to content

Latest commit

 

History

History
98 lines (73 loc) · 4.68 KB

Changelog.md

File metadata and controls

98 lines (73 loc) · 4.68 KB
Raw

Planned changes for v1.0

Features that are planned to be implemented before the release of v1.0.0, in no particular order.

  • Separate Sectors module
  • Make TrivialTensorMap and TensorMap be the same
  • Simplify TensorMap type to hide rowr and colr
  • Change block order in rowr / colr to speed up particular contractions
  • Make AdjointTensorMap generic
  • Rewrite planar operations in order to be AD-compatible
  • Fix rrules for fermionic tensors
  • Fix GPU support
  • Proper threading support
  • Rewrite documentation

Changelog

v0.14

Use DiagonalTensorMap for singular values and eigenvalues

The diagonal (1,1) tensor that contains the singular values or eigenvalues of a tensor are now explicitly represented as DiagonalTensorMap instances.

New index functionality

There are is new functionality for manipulating the spaces associated with a tensor:

  • flip(t, i) changes the duality flag of the ith index of t, in such a way that flipping a pair of contracted indices in an @tensor contraction does not affect the result.
  • insertleftunit(t, i) and insertrightunit(t, i) insert a trivial unit space to the left or to right of index i, whereas removeunit(t, i) removes such a trivial unit space.

SVD truncation change (breaking)

There is a subtle but breaking change in the truncation mechanism in SVD, where now it is guaranteed that smaller singular values are removed first, irrespective of the (quantum) dimension of the sector to which they belong

DiagonalTensorMap and reduceddim

This adds a DiagonalTensorMap type for representing tensor maps in which all of the blocks are diagonal. This only makes sense for maps between a single index in the domain and the codomain (which are furthermore required to be the same), as otherwise the fusion and splitting trees from the domain and codomain to the blocked sectors would itself be nondiagonal. This new type will be used to capture the singular values and eigenvalues as tensor maps in the corresponding decompositions. The number of free parameters in a DiagonalTensorMap instance on vector space V is equal to reduceddim(V), a new function that sums up the degeneracy dimension dim(V, c) for each of the sectors c in V. This function can be useful by itself and is also exported from the TensorKit module. An instance of DiagonalTensorMap can then be created as DiagonalTensorMap(data, V) where data is a vector of length reduceddim(V).

v0.13

AbstractTensorMap{E,S,N₁,N₂}

This adds the scalar type as a parameter to the AbstractTensorMap type. This is useful in the contexts where different types of tensors are used (AdjointTensor, BraidingTensor, ...), which still have the same scalartype. Additionally, this removes many specializations for methods in order to reduce ambiguity errors.

copy(BraidingTensor)

This PR changes the behaviour of copy as an instantiator for creating a TensorMap from a BraidingTensor. The rationale is that while this does sometimes happen in Julia Base, this is always in the context of lazy wrapper types, for which it makes sense to not copy the parent and then create a new wrapper. BraidingTensor does not wrap anything, so this definition makes less sense.

Refactor tensormap constructors

This PR refactors the constructors for TensorMap to be more in line with Array constructors in Julia. In particular, this means that the default way to create uninitialized tensors is now TensorMap{E}(undef, codomain ← domain), reminiscent of Array{E}(undef, dims). Several convenience constructors are also added: ones, zeros, rand and randn construct tensors when dims is replaced by domain ← codomain.

TensorOperations v5

This PR bumps the compatibility of TensorOperations to v5. This is a breaking change as there are some changes in the API.

TensorKitSectors

This promotes TensorKitSectors to its own package, in order to make the dependencies lighter and to separate the concerns of the two packages.

FusionTree vertices

In order to simplify the FusionTree struct, we removed the usage of anything other than Int as a vertex-label when working with GenericFusion sectors.

TensorStructure

This PR changed the data structure of TensorMap to consist of a single vector of data, where all blocks and subblocks are various views into this object. This entails major simplifications for the outwards-facing interface, as now the TensorMap type requires less parameters, and the fusion-tree structure is more clearly considered as an implementation detail that can be left hidden to the user. Various other improvements and documentation work was also carried out.