Using ITensornetwork for different ansatz and model

[nzy1997]

[HsiaHongKuan]

assign to hongkuan

[HsiaHongKuan]

How the algorithm goes, so first generate a random MPS, then apply the simple update to evolve the state, then use the BP to contract?

[nzy1997]

First version in this branch with weird performance as following.
There are some bug in ITensorNetworks.jl. To make this version of code run correctly, we need to dev ITensorNetworks.jl and modify this function in file lib/ITensorExtension/srcitensor.jl

function map_eigvals(f::Function, A::ITensor, inds...; ishermitian=false, kwargs...)
  return map_diag(f, A)
  if isdiag(A)
    return map_diag(f, A)
  end

  Ul, D, Ur = eigendecomp(A, inds...; ishermitian, kwargs...)

  return Ul * map_diag(f, D) * Ur
end

Parameter setting and test code:

using BPGauge
using Test
using ITensorNetworks

@testset "BPGauge" begin
    unit = 3.0
    r = 3.5
    bd = 1
    Ω_max = 2π * 4
    Δ_max = 3 * Ω_max
    t_max = 1.5
    U = Ω_max * 20
    dt = 0.01
    maxdim = 10
    lattice_size = (3,3)

    lattice = SquareLattice()
    atoms = map(x -> unit .* x, generate_sites(lattice, lattice_size...));
    tn,ind_vec,g = create_zero_state(atoms,r, bd)
    for t in dt:dt:t_max
        Ω, Δ = build_adiabatic_sweep(Ω_max, Δ_max, t_max,t)
        @show t Ω Δ 
        tn = apply_rydberg_hamiltonian(Ω,Δ,U,dt,ind_vec,g,tn;maxdim)
    end
    expc = ITensorNetworks.expect(tn, "Sz"; alg="bp")
    @show  real(expc).> 0.0
    @show expc
end

The results are

real(expc) .> 0.0 = Bool[1, 0, 1, 0, 0, 0, 0, 0, 0]
expc = ComplexF64[0.49376070744112877 + 3.055998254466034e-18im, -0.47958601126882555 + 1.1002377878012125e-18im, 0.495564257751671 - 6.854151237485841e-21im, -0.4758890416559952 - 9.198791089437364e-18im, -0.48070387778902995 + 1.5350995412895235e-18im, -0.4931838482974981 + 3.927796531557539e-18im, -0.04836370328253021 - 1.222673500806524e-17im, -0.4648965565351821 + 8.095890467796483e-18im, -0.4040827011411455 + 9.854544668818026e-17im]
Test Summary: | Total   Time
BPGauge       |     0  10.2s
Test.DefaultTestSet("BPGauge", Any[], 0, false, false, true, 1.7491793617371e9, 1.749179371891245e9, false, "/Users/nizhongyi/jcode/BPGauge.jl/test/itensortoolkit.jl")

[nzy1997]

@HsiaHongKuan can you figure out what is wrong with our first version?

[nzy1997]

I add some tests on the codes. I still can't figure out why the independent set is wrong. I try to add maximum bound dimension to 50. The result stays the same.

real(expc) .> 0.0 = Bool[1, 0, 1, 0, 0, 0, 0, 0, 0]
expc = ComplexF64[0.4764049769464117 - 2.9723383494068207e-18im, -0.2851016282529076 - 7.750603198233625e-17im, 0.4878119092569195 + 1.250318554666261e-17im, -0.4994332871853344 - 9.50373936248287e-20im, -0.4976705945356544 - 4.474866075763957e-18im, -0.44405637069836684 - 1.2910417588753963e-17im, -0.4203083224147581 + 4.013353070156908e-18im, -0.4998572768931988 - 5.911595684668246e-19im, -0.4986287144604898 - 2.059378305189573e-18im]
Test Summary: | Total       Time
BPGauge       |     0  102m50.4s

[HsiaHongKuan]

I set the environment to empty set, and the result a little better? @nzy1997

[nzy1997]

Maybe there is some bug in the environment computation?

[HsiaHongKuan]

The parameters are

[xuanzhaogao]

@HsiaHongKuan In that case I suggest you to try the spin glass model with 2-site interactions first.
Additionally, please paste the code directly instead of using images, example: TensorBFS/OMEinsumContractionOrders.jl#62

[HsiaHongKuan]

@ArrogantGao The spin glass model you mentioned is the paper in wechat group you replicate?

[xuanzhaogao]

@ArrogantGao The spin glass model you mentioned is the paper in wechat group you replicate?

no, that is a classical model. See this: https://en.wikipedia.org/wiki/Transverse-field_Ising_model

I think this model is simpler.