TensorKit sectors

Loading TensorKit makes SU(2) at a level, or at real q, available as a tensor symmetry. SU2k(k) and SU2q(q) return sector types whose F- and R-symbols come from fsymbol and rmatrix. The extension needs only TensorKitSectors, so Nemo is not involved.

using TensorKit, QRecoupling

I = SU2k(3)                                  # SU(2) at level 3, q = e^{iπ/5}
a = I(1)                                     # spin 1, also the τ of the Fibonacci category
collect(a ⊗ a)                               # [I(0), I(1)]: fusion is truncated by the level
dim(a)                                       # 1.618…, the quantum dimension
Fsymbol(a, a, a, a, a, a)                    # −0.618…
Rsymbol(a, a, I(0))                          # a phase

V = Vect[I](0 => 2, 1//2 => 2, 1 => 1)
t = randn(ComplexF64, V ⊗ V ← V ⊗ V)
norm(repartition(t, 1, 3)) ≈ norm(t)         # F-moves are unitary
braid(t, ((2, 1), (3, 4)), (1, 2, 3, 4))     # braiding, with levels for the strands
U, S, Vh = svd_compact(t)

The conventions are TensorKit's, F^{abc}_{d,ef} = (−1)^{a+b+c+d} √([2e+1][2f+1]) {a b e; c d f}, which is fsymbol(a, b, e, c, d, f). The field a.J of a label is 2j.

What each type supports

SU2k(k)SU2q(q), q > 0
Spins0, 1/2, …, k/2every multiple of 1/2
Fusiontruncated by the levelas for SU(2)
F-symbolsreal and orthogonalreal and orthogonal
R-symbolscomplex phasesnone: they are not unitary, which TensorKit assumes
Tensor operationsplanar and braidedplanar only

The F-symbols are Float64 in both, so planar tensors can stay real. TensorKit 0.16 still turns real tensors at a level complex, with zero imaginary parts, and 0.17 keeps them real. SU2q(1.0) agrees with SU2Irrep. Fusion tensors are not provided, since the quantum dimension of a sector is not the size of a matrix block, so convert(Array, t) is unavailable, as for TensorKit's other anyons.

TensorKit warns that tensors with real data might be incompatible with SU2k(k). That matters only for a tensor that is braided, since the R-symbols are complex. Nemo and TensorKit both export dim, so with both loaded write TensorKit.dim, or load Nemo with import Nemo.

Large levels

An F-symbol takes 0.1 to 0.6 µs at levels up to 20000, with relative error below 5 × 10⁻¹⁴ on samples against high-precision values. Above level 32 a graded space stores its multiplicities in a dictionary instead of in its type, so large levels do not slow compilation.

Checks

test/tensorkit_environment.jl runs TensorKit's own sector test suite on levels 1, 3, 10 and 40 and on q = 0.8 and 1.3, every pentagon and hexagon equation at levels 2 and 4, sampled ones up to k = 5000, and tensor operations at a level and at real q. The integer spins of level 3 reproduce the chain spectra of TensorKit's FibonacciAnyon, SU2q(1.0) the F-symbols of SU2Irrep, and chains of spins 1/2 at a level and at real q the spectra of the U_q-invariant XXZ chain. Before the release the same checks ran more widely, on levels 0 to 40, five values of q, products with U1Irrep, FermionParity and Z2Irrep, and XXZ chains of up to eight sites. The ground-state energy densities of infinite chains found with MPSKit's VUMPS lie within 6 × 10⁻⁶ above the Bethe-ansatz values from level 2 to level 1000.