Greedy basis assembly using DMRG

As a follow-up example, we now want to showcase how to compute a reduced basis by means of the density matrix renormalization group (DMRG). To that end, we utilize the ITensors.jl package which, among other things, efficiently implements DMRG. We will see that, while we need to adjust the way we set up the model Hamiltonian as well as our solver, most steps stay the same. Again, we treat the one-dimensional $S=1/2$ XXZ model from the previous example.

Hamiltonians as MPOs

Let us begin by building the XXZ Hamiltonian. Instead of constructing explicit matrices from Kronecker products as we did before, we now use a tensor format called matrix product operators (MPOs) to represent the Hamiltonian.

using ITensors
using ReducedBasis
using Random: seed!
seed!(0);

To build the Hamiltonian terms as MPOs, we make use of the ITensors.OpSum() object that automatically produces a MPO from a string of operators. The affine MPO terms are then stored in an AffineDecomposition as ApproxMPOs which also include possible truncation keyword arguments:

function xxz_chain(sites::IndexSet; kwargs...)
    xy_term   = OpSum()
    zz_term   = OpSum()
    magn_term = OpSum()
    for i in 1:(length(sites) - 1)
        xy_term   += "Sx", i, "Sx", i + 1
        xy_term   += "Sy", i, "Sy", i + 1
        zz_term   += "Sz", i, "Sz", i + 1
        magn_term += "Sz", i
    end
    magn_term += "Sz", length(sites)  # Add last magnetization term
    AffineDecomposition([ApproxMPO(MPO(xy_term, sites), xy_term; kwargs...),
                         ApproxMPO(MPO(zz_term, sites), zz_term; kwargs...),
                         ApproxMPO(MPO(magn_term, sites), magn_term; kwargs...)],
                        μ -> [1.0, μ[1], -μ[2]])
end;

So let us instantiate such a MPO Hamiltonian where we also specify a singular value cutoff, which is passed to the ApproxMPO objects:

L = 12
sites = siteinds("S=1/2", L)
H = xxz_chain(sites; cutoff=1e-9);

Notice that we can now choose a bigger system size (which is still very small here), since the tensor format allows for efficient low rank approximations (hence the cutoff) that buy us a substantial performance advantage when going to larger systems.

Using the DMRG solver for obtaining snapshots

Having created our Hamiltonian in MPO format, we now need a solver that is able to compute ground states from MPOs. The corresponding ground state will also be obtained in a tensor format, namely as a matrix product state (MPS). This is achieved by ITensors.dmrg which is wrapped in the DMRG solver type:

dm = DMRG(; sweeps=default_sweeps(; cutoff_max=1e-9),
          observer=() -> DMRGObserver(; energy_tol=1e-9));

For each solve a new ITensors.DMRGObserver object is created that converges the DMRG iteration up the specified energy_tol. The sweeps argument takes any ITensors.Sweeps object that controls the approximation settings with respect to the DMRG sweeps. While the implemented DMRG solver is capable of also solving degenerate ground states, we here opt for non-degenerate DMRG settings (see the n_states and tol_degeneracy keyword arguments in DMRG) which is the default behavior. (We do this due to a $L+1$-fold degeneracy on the parameter domain, where the degenerate DMRG solver can produce instable results for larger $L$.)

As discussed in the last example, we need a way to orthogonalize the reduced basis. Due to the MPS format that the snapshots will have, we cannot use QR decompositions anymore and resort to a different method, EigenDecomposition, featuring an eigenvalue decomposition of the snapshot overlap matrix that can drop insignificant snapshots that fall below a cutoff:

edcomp = EigenDecomposition(; cutoff=1e-7);

Now with different types for the Hamiltonian, the solver and the orthogonalizer, we call assemble using the greedy strategy and training grid from the last example. However, instead of specifying a tolerance for the maximal error estimate of our basis, we now set a maximal number of performed truth solves via n_truth_max:

Δ = range(-1.0, 2.5; length=40)
hJ = range(0.0, 3.5; length=40)
grid_train = RegularGrid(Δ, hJ)
greedy = Greedy(; estimator=Residual(), n_truth_max=24)

rbres = assemble(H, grid_train, greedy, dm, edcomp);

n      max. err    ‖BᵀB-I‖     time      μ
------------------------------------------------------------
1      NaN         8.88e-16     434ms    [-1.0, 0.0]
2      6.3         2.29e-16     628ms    [2.5, 3.5]
3      3.36        3.91e-16     354ms    [2.5, 0.538]
4      1.52        1.74e-14     566ms    [-0.282, 0.09]
5      1.87        1.91e-13     225ms    [2.5, 2.513]
6      0.934       7.49e-15     216ms    [0.795, 0.0]
7      0.726       6.06e-15     212ms    [0.077, 0.718]
8      0.868       5.91e-15     214ms    [2.5, 3.141]
9      0.723       6.31e-15     231ms    [0.795, 0.538]
10     0.814       1.43e-14     245ms    [2.5, 0.897]
11     0.708       1.12e-14     232ms    [2.051, 2.423]
12     0.538       5.14e-14     246ms    [0.705, 0.897]
13     0.443       3.19e-14     329ms    [-0.103, 0.0]
14     0.274       3.78e-14     314ms    [-0.372, 0.269]
15     0.233       1.16e-13     790ms    [-0.91, 0.0]
16     0.229       8.77e-14     293ms    [-0.462, 0.449]
17     0.113       1.1e-13      339ms    [-0.641, 0.09]
18     0.103       1.44e-13     987ms    [-0.821, 0.09]
19     0.079       9.01e-13     427ms    [1.782, 0.359]
20     0.0748      8.76e-13     428ms    [-0.731, 0.09]
21     0.041       1.05e-11     453ms    [1.692, 1.256]
22     0.0393      2.46e-12     488ms    [-0.641, 0.0]
23     0.0269      3.72e-12     483ms    [0.526, 1.436]
24     0.0933      7.78e-12     485ms    [2.5, 3.41]

The returned basis now has snapshot vectors of ITensors.MPS type, which we have to keep in mind when we want to compress observables. That is to say, the observables have to be constructed as AffineDecompositions with ApproxMPO terms as we did for the Hamiltonian. Again, we want to compute the magnetization so that we can reuse the third term of H:

M    = AffineDecomposition([H.terms[3]], [2 / L])
m, _ = compress(M, rbres.basis);

And at that point, we continue as before since we have arrived at the online phase where we only operate in the low-dimensional RB space, agnostic of the snapshot solver method. We have to make sure, however, to choose matching degeneracy settings for the FullDiagonalization solver in the online phase:

fulldiag = FullDiagonalization(dm);

Then we can define an online grid and compute the magnetization at all grid points, again constructing m at an arbitrary parameter point since its coefficient is parameter-independent:

m_reduced = m()
Δ_online = range(first(Δ), last(Δ); length=100)
hJ_online = range(first(hJ), last(hJ); length=100)
grid_online = RegularGrid(Δ_online, hJ_online)

using Statistics
magnetization = map(grid_online) do μ
    _, φ_rb = solve(rbres.h, rbres.basis.metric, μ, fulldiag)
    mean(u -> abs(dot(u, m_reduced, u)), eachcol(φ_rb))
end;

Plotting the magnetization on a heatmap, we arrive at the following result:

using Plots
hm = heatmap(grid_online.ranges[1], grid_online.ranges[2], magnetization';
             xlabel=raw"$\Delta$", ylabel=raw"$h/J$", title="magnetization",
             colorbar=true, clims=(0.0, 1.0), leg=false)
params = unique(rbres.basis.parameters)
scatter!(hm, [μ[1] for μ in params], [μ[2] for μ in params];
         markershape=:xcross, color=:springgreen, ms=3.0, msw=2.0)

We reproduce the ground-state phase diagram, but this time with more magnetization plateaus (due to increased system size) and we see that the greedy algorithm chose different parameter points to solve using DMRG.