GitHub

ITensor Workflows

EDKit includes a practical bridge between exact-diagonalization style data structures and ITensor-based tensor-network workflows.

This layer is useful when you want to:

  • convert state vectors into MPS form,
  • convert dense operators into ITensor operators,
  • move between ordinary MPS/MPO objects and Pauli-space representations,
  • build TEBD gates from local matrices.

Vector And MPS Conversion

The simplest entry points are vec2mps and mps2vec.

using EDKit, ITensors, ITensorMPS, LinearAlgebra

L = 6
s = siteinds(2, L)
psi = normalize(randn(ComplexF64, 2^L))

psi_mps = vec2mps(psi, s)
psi_back = mps2vec(psi_mps)

mps2vec(psi, B) also supports extracting amplitudes in a symmetry-aware basis B when the MPS already lies in that sector.

Operator Conversion

mat2op and op2mat convert between dense matrices and ITensor operator objects:

using EDKit, ITensors

L = 4
s = siteinds(2, L)
h = spin((1.0, "xx"), (1.0, "yy"), (1.0, "zz"))

O = mat2op(h, s[1], s[2])
h_back = op2mat(O, s[1], s[2])

This is useful when you already have a local matrix from the EDKit side and want to feed it into an ITensor workflow.

Pauli-Space Tools

EDKit also supports operator-space calculations in a Pauli basis.

The main pieces are:

  • pauli and pauli_list for converting between matrices and Pauli coefficients,
  • commutation_mat for the superoperator -i[H, \cdot],
  • dissipation_mat for the dissipator D[L],
  • mps2pmps, pmps2mpo, and mpo2pmpo for Pauli-space tensor-network conversions.

This is particularly useful for operator growth and Lindbladian tensor-network workflows.

TEBD Gates

tebd4 builds fourth-order Trotter gate sequences from a list of local Hamiltonian pieces:

using EDKit, ITensors

L = 5
ps = siteinds("Pauli", L)
h2 = commutation_mat(spin((1.0, "xx"), (1.0, "yy"), (1.0, "zz")))
gates = tebd4(fill(h2, L - 1), ps, 0.05)

The lower-level tebd_n! routine applies an n-site TEBD update directly to an MPS.

When To Use This Layer

Use the ITensor layer when exact diagonalization and tensor networks need to meet. If you only need small-system dense or sparse exact diagonalization, you can stay entirely inside the basis and operator layers.